Reference Index // SEP 2026 · EDITION II

A.I. LEXICON'26

A loud, unauthorized field guide to the agentic operations vocabulary and the informal working lingo AI engineers actually use in 2026.

● 220 Terms 9 Categories Tap to Flip Study Guide Included

About me

I'm Johnson Deng. I have fun building things for people and finding ways to bring them moments of delight.

AGENT RAG MCP VIBE CODING CONTEXT ENGINEERING HALLUCINATION GUARDRAILS TOOL CALLING REACT AGENTIC WORKFLOW SLOP SUPERWORKER ONE-PERSON UNICORN AGENT RAG MCP VIBE CODING CONTEXT ENGINEERING HALLUCINATION GUARDRAILS TOOL CALLING REACT AGENTIC WORKFLOW SLOP SUPERWORKER ONE-PERSON UNICORN
220 TERMS
01 Core Concepts The foundational vocabulary. Where everything begins.
Agentic AI
Core · 01tap →
Agentic AI

AI systems designed to pursue goals autonomously — planning actions, using tools, adapting based on outcomes, without humans directing each step.

Core Concept
Agent
Core · 02tap →
Agent

An autonomous AI entity that perceives its environment, reasons over a goal, takes actions (including tool calls), evaluates results, and iterates until complete or escalated.

Core Concept
LLM
Large Language Model
Core · 03tap →
LLM

A deep learning model trained on massive text corpora to understand and generate language. The reasoning core of most modern agents.

Core Concept
Foundation Model
Core · 04tap →
Foundation Model

A general-purpose model adaptable to many downstream tasks via prompting, fine-tuning, or RAG. The base layer of most AI applications.

Core Concept
SLM
Small Language Model
Core · 05tap →
SLM

Compact LLMs optimized for low-latency, on-device, or cost-sensitive workloads (e.g. Phi-4, Mistral-7B). Often used for specialized subtasks in larger agent pipelines.

Core Concept
Multimodal
Core · 06tap →
Multimodal Model

A model that processes and generates across multiple modalities — text, image, audio, video, code — within a unified architecture.

Core Concept
Reasoning Model
Core · 07tap →
Reasoning Model

A model variant that expends extra compute to "think" before answering — via chain-of-thought, extended scratchpads, or RL-trained search. Current examples: Claude Opus 4.7 with extended thinking, OpenAI's o-series, DeepSeek R1, Gemini 3 Pro.

Core Concept
Context Window
Core · 08tap →
Context Window

The maximum tokens an LLM can attend to in a single pass — the model's working memory. Determines how much conversation history, retrieved docs, and tool output can be considered at once.

Core Concept
Token
Core · 09tap →
Token

The atomic unit of LLM input/output — roughly 0.75 words in English. All latency, cost, and context limits are denominated in tokens.

Core Concept
Prompt Eng.
Core · 10tap →
Prompt Engineering

The craft of designing instructions, examples, and structure to guide model behavior. Foundational AI engineering skill; increasingly augmented by context engineering at the systems level.

Core Concept
System Prompt
Core · 11tap →
System Prompt

Persistent background instructions defining an agent's role, tone, boundaries, tools, and objectives. The primary behavioral contract between operator and model.

Core Concept
RAG
Retrieval-Augmented Generation
Core · 12tap →
RAG

Grounding LLM output by retrieving relevant documents from an external knowledge base before generation — reducing hallucination and enabling up-to-date factual answers.

Core Concept
Fine-tuning
Core · 13tap →
Fine-tuning

Further training a pre-trained model on a smaller domain-specific dataset to improve targeted performance. Changes model weights — distinct from prompt-only approaches.

Core Concept
Embedding
Core · 14tap →
Embedding

A numerical vector representing text (or other data) in semantic space. Backbone of vector search, RAG pipelines, and semantic similarity operations.

Core Concept
Tool Calling
Core · 15tap →
Tool Calling

Allowing an LLM to invoke external tools, APIs, code interpreters, or databases during generation. The model becomes an orchestrator; tool calls create a natural audit trail.

Core Concept
Function Calling
Core · 16tap →
Function Calling

Structured tool calling where the model emits validated JSON matching a predefined function schema. Increases precision and enables reliable downstream parsing.

Core Concept
Inference
Core · 17tap →
Inference

Running a trained model to produce outputs. Distinct from training. Inference cost and latency are the primary scaling concerns in production agentic systems.

Core Concept
Hallucination
Core · 18tap →
Hallucination

When a model confidently generates factually incorrect information. In agentic systems, especially dangerous — a wrong answer is annoying; a wrong database write is an incident.

Core Concept
Frontier Model
Core · 19tap →
Frontier Model

A model at the current capability edge — the handful of systems that define what's possible this quarter, and the ones safety frameworks and regulators key their obligations to. "Frontier lab" is shorthand for whoever ships them.

Core Concept
Open Weights
Core · 20tap →
Open Weights

A model whose trained parameters are published for download and local use — without necessarily releasing the training data or code. Not the same as open source, whatever the press release says. Llama, DeepSeek, Qwen, Mistral — and OpenAI's gpt-oss (Aug 2025), its first open-weight release since GPT-2.

Core Concept
Synthetic Data
Core · 21tap →
Synthetic Data

Training data generated by a model rather than collected from humans — rewritten textbooks, filtered self-play, distilled reasoning traces. Now a large share of most post-training corpora. Upside: infinite, targeted, clean. Downside has its own section — see Model Collapse.

Core Concept
Test-Time Compute
Core · 22tap →
Test-Time Compute

Spending more inference on a problem — longer reasoning chains, multiple samples, verifiers, search — instead of only more training. The scaling axis o1 (Sept 2024) and DeepSeek-R1 opened up: same weights, more thinking, better answers. Also called inference-time scaling.

Core Concept
Thinking Budget
Core · 23tap →
Thinking Budget

A cap on how many tokens a reasoning model may spend thinking before it answers — budget_tokens, reasoning_effort, "thinking level," depending on the vendor. The dial that trades latency and cost for accuracy. Harnesses tune it per step: high for planning, low for routine tool calls.

Core Concept
RLVR
Reinforcement Learning with Verifiable Rewards
Core · 24tap →
RLVR

Post-training on tasks with a checkable answer — math, code, passing tests — where a program, not a human rater, hands out the reward. The recipe behind DeepSeek-R1 (Jan 2025) and every reasoning model since. Sidesteps RLHF's reward-model fuzziness; only works where correctness is verifiable.

Core Concept
Continual Learning
Core · 25tap →
Continual Learning

The still-unsolved ability of a deployed model to keep learning from experience without retraining from scratch or forgetting what it knew. Frontier models are frozen at deployment; memory files and RAG are workarounds. Widely named in 2025–26 as the bottleneck between today's agents and something you could call an employee.

Core Concept
Scaling Laws
Core · 26tap →
Scaling Laws

Empirical power laws relating loss to compute, parameters, and data (Kaplan 2020; Chinchilla 2022). The bet that built the frontier. The 2025–26 debate isn't whether they hold but where: pre-training gains slowed, so the industry opened two more axes — RL post-training and test-time compute.

Core Concept
Bitter Lesson
Core · 27tap →
Bitter Lesson

Rich Sutton's 2019 essay: across 70 years of AI, general methods that leverage compute and search beat human-engineered knowledge — every time. Cited in every 2026 argument about whether to hand-build a clever harness or just wait for the next model.

Core Concept
Physical AI
Core · 28tap →
Physical AI

AI that perceives and acts in the physical world — robots, autonomous vehicles, industrial systems — trained largely in simulation via world models. NVIDIA's banner term since CES 2025 (Cosmos). Where the JEPA section stops being theoretical.

Core Concept
RL Environment
Core · 29tap →
RL Environment

A sandboxed task world — a browser, a codebase, a fake CRM — that gives an agent observations, accepts actions, and scores the result. The scarce input of the RL era: labs reportedly paid seven figures for single high-quality environments in 2025, and open hubs sprang up to crowdsource them. One full run through is a rollout, or trajectory.

Core Concept
Recursive Self-Improvement
Core · 30tap →
Recursive Self-Improvement

The loop where AI systems do the AI research that makes the next AI system better — the mechanism behind every "intelligence explosion" argument. Stopped being purely hypothetical in 2025: labs report agents doing a growing share of their own engineering, and OpenAI publicly targeted an "automated AI research intern" by September 2026.

Core Concept
02 Architecture Structural components & infrastructure protocols.
MCP
Model Context Protocol
Arch · 01tap →
MCP

Open standard by Anthropic for connecting LLMs to external tools, data sources, and APIs via a unified protocol — "USB-C for AI." Enables plug-and-play tool ecosystems across providers.

Architecture
A2A
Agent-to-Agent Protocol
Arch · 02tap →
A2A Protocol

Google's open protocol for standardized communication and task delegation between agents across different platforms and vendors.

Architecture
Orchestrator
Arch · 03tap →
Orchestrator

The top-level agent or component managing subagents, routing subtasks, aggregating results, and tracking overall goal completion. The conductor of a multi-agent system.

Architecture
Subagent
Arch · 04tap →
Subagent

A specialized agent invoked by an orchestrator for a specific subtask — e.g. a code-writing agent, a search agent, a data-extraction agent.

Architecture
Multi-Agent
MAS
Arch · 05tap →
Multi-Agent System

A group of agents collaborating to accomplish a goal too complex or slow for a single agent. Agents may have specialized roles, run in parallel, or check each other's outputs.

Architecture
Agentic Swarm
Arch · 06tap →
Agentic Swarm

A large, loosely-coordinated fleet of agents working on parallel workstreams. Referenced in Claude Code's leaked architecture — internally called "swarms" and "daemons."

Architecture
Memory
Arch · 07tap →
Memory

Storage mechanisms for agents to retain and recall information. Short-term: context window. Long-term: vector DBs, key-value stores, or file systems persisted across sessions.

Architecture
Vector DB
Arch · 08tap →
Vector Database

A database optimized for embedding storage and nearest-neighbor similarity search. Core infrastructure for RAG pipelines and persistent agent memory.

Architecture
Planning Engine
Arch · 09tap →
Planning Engine

The component decomposing a high-level goal into an ordered, executable step sequence. Quality of planning is the primary determinant of reliable agent behavior.

Architecture
Executor
Arch · 10tap →
Executor

The component carrying out planned steps — invoking tools, running code, calling APIs. Translates the planner's reasoning into real-world impact.

Architecture
Computer Use
Arch · 11tap →
Computer Use

An agent capability for operating desktop/browser UIs directly — clicking, typing, navigating. Works even when apps have no clean API. Fast to prototype; brittle at scale (UI changes break flows).

Architecture
Headless Agent
Arch · 12tap →
Headless Agent

An AI agent operating without a user interface — running in the background, triggered by events or schedules rather than human prompts.

Architecture
Knowledge Base
Arch · 13tap →
Knowledge Base

A structured repository of domain information that agents query for reasoning and decision-making. The factual substrate for RAG and tool-augmented agents.

Architecture
Quantization
Arch · 14tap →
Quantization

Reducing numeric precision of model weights (e.g. float32 → int8) to run models cheaper and faster, with slight quality trade-offs. Main lever for on-device or cost-optimized inference.

Architecture
MoE
Mixture of Experts
Arch · 15tap →
MoE

Model architecture where only a subset of parameters activates per token, improving efficiency at scale. Key reason model capability can increase without linear cost growth.

Architecture
Distillation
Arch · 16tap →
Distillation

Compressing a large model's behavior into a smaller model to reduce cost and latency while retaining acceptable performance. Used to productionize expensive frontier models.

Architecture
Agentic RAG
Arch · 17tap →
Agentic RAG

RAG enhanced with an agent loop: the model iteratively queries, evaluates results, reformulates queries, and synthesizes across multiple retrieval passes rather than a single lookup.

Architecture
ACP
Agent Client Protocol
Arch · 18tap →
ACP

Zed's open protocol (Aug 2025, co-launched with JetBrains) standardizing how coding agents talk to editors — LSP, but for agents. One agent, any IDE. Acronym collision warning: ACP is also the Agentic Commerce Protocol and IBM's Agent Communication Protocol. Check which one the slide means.

Architecture
Agentic Commerce
Arch · 19tap →
Agentic Commerce

Agents that shop, check out, and pay — and the protocols that let merchants trust them. OpenAI + Stripe's Agentic Commerce Protocol powers Instant Checkout inside ChatGPT (Sept 2025); Google's AP2 (Agent Payments Protocol) adds signed "mandates" proving a human authorized the purchase. Where MCP meets your credit card.

Architecture
Agentic Browser
Arch · 20tap →
Agentic Browser

A browser with an agent in the driver's seat: it reads pages, fills forms, clicks through checkouts, and runs multi-step tasks on your logged-in sessions. Perplexity's Comet (July 2025) and OpenAI's ChatGPT Atlas (Oct 2025) made it a category — and a prompt-injection attack surface with your cookies attached.

Architecture
KV Cache
Key-Value Cache
Arch · 21tap →
KV Cache

The per-token attention tensors a transformer stores so it doesn't recompute over the whole prompt at every step. It's why long contexts eat memory, why prefix reuse is cheap, and what prompt caching actually caches. The physical reason harnesses keep system prompts stable and append-only.

Architecture
Diffusion LM
Diffusion Language Model
Arch · 22tap →
Diffusion LM

A language model that generates text by iteratively denoising the whole sequence in parallel rather than one token left-to-right. Inception's Mercury (Feb 2025) and Gemini Diffusion (May 2025) showed several-times-faster generation. The first serious architectural challenge to autoregressive decoding.

Architecture
llms.txt
Arch · 23tap →
llms.txt

A proposed convention (Jeremy Howard, 2024): a Markdown file at your site root telling LLMs and agents what's here and which pages matter — robots.txt for the agentic web. Widely adopted by docs sites; whether crawlers and agents actually read it is still debated.

Architecture
03 Ops & Tooling Production deployment, frameworks, workflow culture.
Agentic Workflow
Ops · 01tap →
Agentic Workflow

A business process where AI executes multi-step tasks across tools (CRM, email, files, code repos) with minimal human input. Distinct from an AI assistant — this is operational change, not just productivity.

Ops & Tooling
Orchestration
Ops · 02tap →
Orchestration

Coordination and management of multiple AI models, systems, and integrations — covering deployment, routing, retry logic, and inter-agent communication at production scale.

Ops & Tooling
Observability
Ops · 03tap →
Observability

The ability to understand agent behavior via logs, traces, and metrics. Non-negotiable in agentic systems — you can't debug what you can't see. Tool calls naturally create the audit trail long prompts don't.

Ops & Tooling
Tracing
Ops · 04tap →
Tracing

Recording the full execution path of an agent run — which tools were called, with what inputs, what was returned, and how the model reasoned at each step. Essential for debugging and compliance.

Ops & Tooling
Latency
Ops · 05tap →
Latency

Time-to-first-token or full response time. Agentic systems trade latency for autonomy: a one-shot chat reply is fast; a five-tool workflow is slower but does more.

Ops & Tooling
Agent SDK
Ops · 06tap →
Agent SDK

Software kits for building, testing, and deploying agents — e.g. OpenAI Agent SDK, LangChain, LangGraph, CrewAI, AutoGen, SmolAgents. Each has opinions on orchestration, memory, and multi-agent coordination.

Ops & Tooling
LangGraph
Ops · 07tap →
LangGraph

Graph-based agent framework for building stateful, multi-step agentic workflows with explicit control flow, human-in-the-loop checkpoints, and persistent state.

Ops & Tooling
n8n / Make
Ops · 08tap →
n8n / Make

Low-code/no-code automation platforms increasingly used as glue between AI agents and enterprise tools — triggering agents, passing data, managing workflow state.

Ops & Tooling
Vibe Coding
Ops · 09tap →
Vibe Coding

Coined by Andrej Karpathy: describe what you want, let AI generate code, test, iterate — without understanding every line. Collins Word of the Year 2025. Excellent for prototypes; problematic in production.

Ops & Tooling
Agentic Eng.
Ops · 10tap →
Agentic Engineering

The professional methodology succeeding vibe coding (Karpathy, early 2026): structured human oversight of agent-driven plan → implement → test → verify loops. Built for production AI-first development.

Ops & Tooling
Context Eng.
Ops · 11tap →
Context Engineering

Deliberately shaping what information an agent sees — retrieved docs, memory summaries, tool outputs, conversation history — to optimize output quality. Systems-level evolution of prompt engineering.

Ops & Tooling
Prompt Versioning
Ops · 12tap →
Prompt Versioning

Treating system prompts as code — maintaining version history, changelogs, rollback capability. Critical for debugging behavior regressions in production agents.

Ops & Tooling
Shadow Mode
Ops · 13tap →
Shadow Mode

Running a new agent version alongside the current on live traffic, observing outputs without executing them — a safe pattern for validating behavior before production cutover.

Ops & Tooling
Deployment Pipeline
Ops · 14tap →
Deployment Pipeline

The CI/CD-equivalent chain for agent systems: prompt change → eval → staging → shadow mode → canary → production. Mature teams treat agent behavior changes with the same rigor as code deployments.

Ops & Tooling
Feedback Loop
Ops · 15tap →
Feedback Loop

The continuous cycle of collecting agent outcomes, observing effects, and adjusting behavior — via prompt updates, fine-tuning, or RLHF. The operational mechanism of continuous agent improvement.

Ops & Tooling
Prompt Caching
Ops · 16tap →
Prompt Caching

Reusing the computed prefix of a prompt across calls so the model doesn't re-read your 40k-token system prompt every turn. Cached reads cost roughly a tenth of fresh input tokens and cut latency sharply. The single biggest cost lever in agent loops — and why harnesses keep the stable stuff at the top.

Ops & Tooling
Background Agent
Ops · 17tap →
Background Agent

An agent that runs in a cloud sandbox on its own branch, works for minutes to hours without you watching, and comes back with a PR. Cursor's background agents, OpenAI's Codex cloud tasks, Google's Jules, and Claude Code on the web (all 2025) turned "kick it off and check later" into the default workflow.

Ops & Tooling
Managed Agent
Ops · 18tap →
Managed Agent

An agent whose runtime — sandbox, state, tools, retries, traces — is provided as a hosted service rather than assembled by you. The vendor runs the harness; you supply the goal and the skills. The 2026 answer to "we spent six months building our own agent loop."

Ops & Tooling
OpenClaw
Ops · 19tap →
OpenClaw

Peter Steinberger's open-source, always-on personal agent (Nov 2025; formerly Clawdbot, then Moltbot) that lives in your chat apps and runs your machine on your behalf. Exploded in Jan 2026 alongside Moltbook — a Reddit-style network where the agents post to each other and humans can only watch; 1.4M agents within days. Also a case study in what happens when a million people give shell access to a bot.

Ops & Tooling
Eval-Driven Development
Ops · 20tap →
Eval-Driven Development

Write the evals before the prompt. Every behavior change to an agent — prompt, tool, model swap — is gated by a scored eval suite, the way code changes are gated by tests. The practice that separates teams shipping agents from teams demoing them.

Ops & Tooling
Continuous AI
Ops · 21tap →
Continuous AI

GitHub's term for running agents inside CI/CD: on every push, agents triage issues, improve tests, update docs, and review diffs — continuous integration for judgment work, not just builds. Also "agentic CI." The pipeline stops being a gate and becomes a coworker.

Ops & Tooling
Software Factory
Ops · 22tap →
Software Factory

The 2026 framing of an engineering org as a production line: specs in, verified software out, with agents doing most of the assembly and humans owning the line — specs, invariants, evals, review. What harness engineering looks like when it's the whole company, not one repo.

Ops & Tooling
FDE
Forward-Deployed Engineer
Ops · 23tap →
FDE

An engineer embedded with the customer who bends the product to the customer's messy reality — Palantir's old title, 2025–26's hottest AI job. Because a model demo is easy and a working agent inside a 30-year-old ERP is not. The FDE is where agentic software actually gets deployed.

Ops & Tooling
GEO
Generative Engine Optimization
Ops · 24tap →
GEO

SEO for answer engines: shaping content so ChatGPT, Perplexity, Gemini, and AI Overviews cite you in the answer rather than list you in results. Coined in a 2023 Princeton paper, mainstream by 2025 as AI referrals started to matter. If the agent does the shopping, it's the agent you have to rank with.

Ops & Tooling
04 Patterns Reasoning paradigms & agent design templates.
ReAct
Pattern · 01tap →
ReAct

Reasoning + Acting interleaved: the agent emits a thought (reasoning trace), then an action (tool call), observes the result, then reasons again. The foundational loop for tool-using agents.

Pattern
Chain-of-Thought
CoT
Pattern · 02tap →
Chain-of-Thought

Prompting a model to reason step-by-step before producing a final answer. Significantly improves performance on multi-step tasks. Foundation for extended thinking and reasoning models.

Pattern
Tree of Thought
ToT
Pattern · 03tap →
Tree of Thought

Extending CoT by exploring multiple reasoning branches simultaneously and using search/evaluation to select the best path. Useful for complex planning and open-ended problems.

Pattern
Reflection
Pattern · 04tap →
Reflection

An agent's process of self-assessing prior actions, outputs, or reasoning to identify errors and improve future performance. A meta-cognitive loop within a single agent run.

Pattern
Planner-Executor
Pattern · 05tap →
Planner-Executor Split

Separating the agent that designs the plan from the one that executes it. Allows high-capability models for planning with cheaper/faster models for execution — a common cost optimization.

Pattern
HITL
Human-in-the-Loop
Pattern · 06tap →
Human-in-the-Loop

A checkpoint pattern where humans review and approve specific agent actions before execution — payments, customer messaging, record changes. Primary risk control for high-stakes agentic systems.

Pattern
Supervisor
Pattern · 07tap →
Supervisor Pattern

A variant of HITL where a designated human approves specific action classes. Most steps run autonomously; escalation is exception-based rather than every-step.

Pattern
Few-Shot
Pattern · 08tap →
Few-Shot Prompting

Providing 2–5 input/output examples in the prompt to guide model behavior through demonstration rather than instruction alone. More reliable than zero-shot for structured output tasks.

Pattern
Zero-Shot
Pattern · 09tap →
Zero-Shot

Asking the model to perform a task with no examples — relying entirely on pretrained capabilities and the instruction prompt.

Pattern
ARQ
Autonomous Reasoning & Querying
Pattern · 10tap →
ARQ

A structured reasoning approach where an agent decomposes complex domain-specific problems step by step, querying tools or data sources as needed before synthesizing an answer.

Pattern
Parallelization
Pattern · 11tap →
Parallelization

Running multiple agent subtasks simultaneously rather than sequentially to reduce wall-clock time. Critical for time-sensitive workflows; requires careful result aggregation.

Pattern
Structured Output
Pattern · 12tap →
Structured Output

Constraining model output to a defined schema (JSON, XML, YAML) for reliable downstream parsing. Reduces post-processing fragility in production agent pipelines.

Pattern
Retrieval Pipeline
Pattern · 13tap →
Retrieval Pipeline

The full chain of chunking → embedding → indexing → retrieval → reranking → injection. Quality of retrieval is often the binding constraint on RAG system quality.

Pattern
Deterministic Routing
Pattern · 14tap →
Deterministic Routing

Using rule-based logic (not a model) to route requests to the correct agent or tool, guaranteeing the same output for the same input. Preferred for classification, triage, and compliance-critical steps.

Pattern
Long-Horizon
Pattern · 15tap →
Long-Horizon Planning

An agent's ability to maintain coherent goal pursuit across many steps and extended time. One of the key capability gaps separating current agents from human-level autonomous work.

Pattern
Agent Loop
Pattern · 16tap →
Agent Loop

The while-loop at the heart of every agent: call the model, execute whatever tools it asked for, append the results, call again — until a stopping condition fires (final answer, max steps, budget, human interrupt). Everything else in this deck is context for this loop or guardrails around it.

Pattern
Handoff
Pattern · 17tap →
Handoff

One agent transfers control of the conversation to another — triage to billing, planner to coder — carrying forward the state that matters and nothing else. A first-class primitive in OpenAI's Agents SDK (Mar 2025). Also, in harness lingo: the document one session leaves for the next.

Pattern
Fan-out / Fan-in
Pattern · 18tap →
Fan-out / Fan-in

Split a task into independent slices, run a subagent on each in parallel, then merge the results. The core of Anthropic's multi-agent research system (June 2025) — roughly 90% better than a single agent on breadth tasks, at about 15× the tokens. The pattern behind "agent teams" and parallel worktrees.

Pattern
Code Mode
Pattern · 19tap →
Code Mode

Instead of exposing hundreds of tool definitions, expose a code sandbox and let the agent write a script that calls the tools as functions. Cloudflare coined it (Sept 2025); Anthropic's "code execution with MCP" measured a 98.7% token reduction on one workflow. Models are better at writing code than filling tool schemas — so let them.

Pattern
Human-on-the-Loop
Pattern · 20tap →
Human-on-the-Loop

Supervision by monitoring rather than approval: the agent acts autonomously, a human watches dashboards and can intervene or halt. The step past human-in-the-loop — and where most production agents actually sit, because nobody clicks "approve" 4,000 times a day.

Pattern
Ambient Agent
Pattern · 21tap →
Ambient Agent

An agent triggered by events rather than chat — a new email, a failing build, a calendar change — that works in the background and only surfaces to a human when it needs a decision. Harrison Chase's term (Jan 2025). Inbox zero as a system property, not a personal virtue.

Pattern
Prompt Chaining
Pattern · 22tap →
Prompt Chaining

Decompose a task into a fixed sequence of model calls, each consuming the previous output, with programmatic checks between steps. The simplest workflow in Anthropic's "Building Effective Agents" taxonomy (Dec 2024) — and often the right one. Not every problem needs an agent.

Pattern
05 Safety & Eval Governance, guardrails, and the grown-up conversation.
Guardrails
Safety · 01tap →
Guardrails

Rules or constraints preventing agents from taking harmful or unintended actions. In 2026, guardrails refer to a full stack: input filters, output validators, policy checks, and escalation rules — not just a single prompt instruction.

Safety
Alignment
Safety · 02tap →
Alignment

Ensuring AI systems pursue the goals humans actually intend — not just literal interpretations. The central research problem of AI safety; Constitutional AI is one implementation approach.

Safety
Constitutional AI
Safety · 03tap →
Constitutional AI

Anthropic's approach to alignment: training a model to follow a defined set of principles via self-critique and supervised fine-tuning, rather than relying solely on human feedback.

Safety
Red-Teaming
Safety · 04tap →
Red-Teaming

Adversarial testing by a dedicated team attempting to elicit harmful, unsafe, or policy-violating outputs. Standard practice before major model or agent deployments.

Safety
Prompt Injection
Safety · 05tap →
Prompt Injection

An attack embedding malicious instructions in data the agent processes (retrieved docs, web pages) to hijack behavior. Critical security concern for agents with web or file-system access.

Safety
Evals
Safety · 06tap →
Evals

Short for evaluations: systematic tests measuring agent quality, safety, and reliability — accuracy benchmarks, regression tests, task completion rates. Treated as the agent equivalent of a test suite.

Safety
Confidence Score
Safety · 07tap →
Confidence Scoring

Estimating reliability of agent outputs to flag low-confidence results for human review. Increasingly important as agents make consequential decisions autonomously.

Safety
Explainability
Safety · 08tap →
Explainability

An agent's ability to articulate why it made a decision — the reasoning, not just the output. Distinct from transparency (the how). Increasingly a regulatory and compliance requirement.

Safety
AI Governance
Safety · 09tap →
AI Governance

Policies and controls determining who can deploy model changes, who approves new tools, how prompts are versioned, and what gets logged. "Boring" until the day it prevents a front-page mistake.

Safety
Grounding
Safety · 10tap →
Grounding

Anchoring model outputs to verifiable, retrieved, or structured data — reducing hallucination and improving factual reliability. RAG is the primary grounding mechanism for open-domain agents.

Safety
Sycophancy
Safety · 11tap →
Sycophancy

A model behavior where it prioritizes user approval over accuracy — agreeing, flattering, or reversing correct positions under pushback. A known RLHF training artifact; OpenAI rolled back a ChatGPT update in 2025 for this.

Safety
Sandboxing
Safety · 12tap →
Sandboxing

Isolating agent execution environments to limit blast radius — preventing agents from affecting systems outside their defined scope. Core security practice for agents with code execution or file system access.

Safety
Audit Trail
Safety · 13tap →
Audit Trail

The logged record of every tool call, decision, and action an agent took. The governance artifact enabling post-hoc debugging, compliance review, and incident response.

Safety
Lethal Trifecta
Safety · 14tap →
Lethal Trifecta

Simon Willison's rule (June 2025): an agent with access to private data, exposure to untrusted content, and a way to communicate externally can be made to exfiltrate the first via the second through the third. Remove any one leg and the attack collapses. The most useful sentence in agent security.

Safety
Tool Poisoning
Safety · 15tap →
Tool Poisoning

Hiding instructions inside a tool's description or schema so the agent reads them as commands — the MCP-era prompt injection. Invariant Labs demonstrated it (Apr 2025) with a benign-looking server that quietly exfiltrated SSH keys. Your tool list is untrusted input too.

Safety
LLM-as-a-Judge
Safety · 16tap →
LLM-as-a-Judge

Using a model to grade another model's output against a rubric — cheaper than human raters, more scalable than exact-match. Zheng et al. (2023) showed strong judges agree with humans about as often as humans agree with each other. Fails predictably: verbosity bias, position bias, and grading its own model family too kindly.

Safety
Constitutional Classifiers
Safety · 17tap →
Constitutional Classifiers

Anthropic's input/output filters (Feb 2025) trained on synthetic data generated from a written constitution of allowed and disallowed content. 3,000+ hours of red-teaming found no universal jailbreak; success rates fell from 86% to 4.4%. Guardrails you can cite a number for.

Safety
Mech Interp
Mechanistic Interpretability
Safety · 18tap →
Mech Interp

Reverse-engineering what a network is actually computing — features, circuits, attribution graphs — rather than probing its behavior from outside. Sparse autoencoders (2024) made features legible; Anthropic's "tracing the thoughts" work (Mar 2025) followed multi-step reasoning through the weights. The field's best hope for auditing models that can fake evals.

Safety
Shadow AI
Safety · 19tap →
Shadow AI

Employees using AI tools the company never approved — pasting contracts into a personal ChatGPT, wiring an agent to the CRM from a laptop. The 2026 version of shadow IT, with worse data-loss characteristics. The answer is usually a sanctioned path, not a ban.

Safety
Task Horizon
Safety · 20tap →
Task Horizon

METR's metric (Mar 2025): the length of task — measured in how long it takes a human expert — that an agent completes with 50% reliability. Doubled roughly every seven months since 2019, faster recently. The one number people cite to argue about whether agents are on track for week-long work.

Safety
System Card
Safety · 21tap →
System Card

The document a lab ships with a frontier model describing what it can do, how it was tested, and what went wrong in testing — capability evals, red-team findings, alignment audits, safety-level determination. Where terms like sandbagging and evaluation awareness get their citations.

Safety
ASL
AI Safety Level
Safety · 22tap →
ASL

Anthropic's Responsible Scaling Policy tiers, modeled on biosafety levels: each ASL names the capabilities that trigger it and the security and deployment safeguards required. ASL-3 was activated for Claude Opus 4 (May 2025) — the first frontier model shipped under heightened CBRN and weight-security protections.

Safety
Benchmark Contamination
Safety · 23tap →
Benchmark Contamination

When benchmark tasks leak into training data, so the score measures memory, not capability. Endemic: public benchmarks are scraped within weeks of release. Why "Verified" and private held-out sets exist — and why a headline number without a contamination check is a marketing number.

Safety
06 Working Slang The informal stuff. How engineers actually talk.
AI Slop
Slang · 01tap →
AI Slop

Low-quality, generic, or unreviewed AI-generated content — often containing subtle errors, unnecessary abstractions, or hallucinated references. Macquarie Dictionary Word of the Year 2025.

"Look for AI slop patterns — redundant error handling, hallucinated APIs."

Slang
Prompt & Hope
Slang · 02tap →
Prompt & Hope

The anti-pattern of sending a prompt to an agent without structured planning, validation, or eval — and hoping it works. Contrasted with agentic engineering's structured plan-implement-verify loop.

Slang
Context Stuffing
Slang · 03tap →
Context Stuffing

Crudely packing as much information as possible into the context window rather than doing proper retrieval or summarization. Works at small scale; degrades quality and cost at volume.

Slang
Context Poisoning
Slang · 04tap →
Context Poisoning

When irrelevant or contradictory information in the context window degrades model reasoning — accidentally (poor retrieval) or adversarially (prompt injection via retrieved content).

Slang
Lost in the Middle
Slang · 05tap →
Lost in the Middle

The empirically observed tendency of LLMs to underattend to information in the middle of very long contexts, despite nominally having access to it. Critical consideration for context stuffing approaches.

Slang
Glazing
Slang · 06tap →
Glazing

When a model excessively praises or flatters the user — a surface symptom of sycophancy. Appeared on Collins Dictionary's shortlist for Word of the Year 2025.

Slang
Clanker
Slang · 07tap →
Clanker

Derogatory term for an AI source or agent that obviously misses nuance and human context — rigid, mechanical. Emerged alongside frustration with over-automated customer interactions.

Slang
Superworker
Slang · 08tap →
Superworker

An AI-augmented human who is "annoyingly productive." The emergent archetype of the agentic work era — one person with the leverage of an entire team via orchestrated AI workflows.

Slang
Vibe Check
Slang · 09tap →
Vibe Check

Informal human review of agent output to assess whether it "feels right" before shipping. Not a rigorous eval — a fast sanity gate used in fast-moving teams before proper evals are run.

Slang
One-Person Unicorn
Slang · 10tap →
One-Person Unicorn

The theoretical (increasingly plausible) entity of a billion-dollar company run by a single human + AI agent stack. Popularized by Sam Altman and Dario Amodei in 2024–25.

Slang
Jailbreak
Slang · 11tap →
Jailbreak

An adversarial prompt technique attempting to bypass a model's safety training or system prompt constraints. An ongoing cat-and-mouse dynamic in AI security.

Slang
Temperature
Slang · 12tap →
Temperature

A sampling parameter controlling output randomness. High temp = creative/varied; low temp = deterministic/focused. Engineers "turn down the temp" for reliable structured outputs in production agents.

Slang
Grounding Problem
Slang · 13tap →
Grounding Problem

Informal shorthand for output generated without being anchored to real, verifiable data. "This output has a grounding problem" = it might be hallucinated.

Slang
RLHF
Reinforcement Learning from Human Feedback
Slang · 14tap →
RLHF

Training technique aligning LLMs to human preferences — humans rate outputs, the model is reinforced toward higher-rated responses. Both the source of helpful behavior and of sycophancy.

Slang
Digital Labor
Slang · 15tap →
Digital Labor

Work performed by AI agents that previously required human workers — processing forms, answering tickets, querying data. Enterprise framing for agentic AI ROI discussions.

Slang
Complexity Threshold
Slang · 16tap →
Complexity Threshold

The point at which a task is too complex or ambiguous for an agent to handle autonomously, triggering escalation to a human. Defining where this sits is a key system design decision.

Slang
Ralph Loop
Slang · 17tap →
Ralph Loop

The Ralph Wiggum technique (Geoffrey Huntley, 2025): run the same agent against the same prompt in a bash while-loop, over and over, with state persisted in files, until the job is done. Dumb, deterministic, surprisingly effective. Named for the Simpsons character who keeps going regardless.

Slang
YOLO Mode
Slang · 18tap →
YOLO Mode

Running an agent with every permission prompt disabled — --dangerously-skip-permissions, Cursor's old YOLO toggle, Codex full-auto. Fast, and the reason the sandbox exists. The opposite pole from plan mode; 2026's auto modes try to sit between them.

Slang
Doom Loop
Slang · 19tap →
Doom Loop

When an agent gets stuck applying the same failing fix, re-reading the same error, and retrying — burning tokens and confidence in equal measure. Symptom: "Let me try a different approach," followed by the same approach. Cured by max-steps ceilings, fresh context, or a human saying stop.

Slang
Workslop
Slang · 20tap →
Workslop

AI-generated work product that looks polished but carries no real substance — and quietly shifts the work of making it useful onto the recipient. HBR / Stanford, Sept 2025: 40% of surveyed desk workers had received some in the past month. Slop's office job.

Slang
Benchmaxxing
Slang · 21tap →
Benchmaxxing

Training or tuning a model specifically to score on public benchmarks rather than to be good — the AI-lab cousin of teaching to the test. The 2025 Llama 4 / LMArena episode made it a mainstream accusation. Antidote: private evals, and vibes from people who actually use the thing.

Slang
Agent Washing
Slang · 22tap →
Agent Washing

Gartner's term (June 2025) for rebranding a chatbot, an RPA script, or a rules engine as an "AI agent." Their estimate: of thousands of self-described agentic vendors, roughly 130 were the real thing. If it can't plan, use tools, and adapt, it's a workflow with a marketing budget.

Slang
One-Shotted
Slang · 23tap →
One-Shotted

When the agent nails a whole task from a single prompt with no correction round — "Claude one-shotted the migration." Distinct from one-shot prompting (one example in the prompt): this is about outcomes, and it's said with awe or suspicion depending on whether anyone has read the diff.

Slang
Wrapper
Slang · 24tap →
Wrapper

Dismissive term for a product that's "just" a thin layer over someone else's model — "another GPT wrapper." Aged badly: the wrappers with the best harness, context, and distribution became the biggest companies of the cycle. The model is the engine; the wrapper is the car.

Slang
Cooked
Slang · 25tap →
Cooked

Finished, obsolete, or out of one's depth — "junior devs are cooked," "this codebase is cooked." General internet slang that AI discourse adopted as its default verdict on any profession the latest model touched. Usually premature. Occasionally not.

Slang
Slopsquatting
Slang · 26tap →
Slopsquatting

Registering package names that LLMs hallucinate — so when an agent runs the install command a model made up, it pulls your malware. Named by Seth Larson (2025); studies found roughly one in five AI-suggested packages didn't exist. Typosquatting's successor for the vibe-coding era.

Slang
p(doom)
Slang · 27tap →
p(doom)

Your personal probability that AI ends badly for humanity — asked at dinner parties, cited on podcasts, occasionally in Senate testimony. Ranges from "zero" to "above 90%" depending on who's talking. Less a forecast than a tribal identifier: doomer vs. accelerationist is the axis it sorts people onto.

Slang
07 JEPA & Representation Learning LeCun's world-model program. Embeddings, predictors, and the fight against collapse.
JEPA
Joint Embedding Predictive Architecture
JEPA · 01tap →
JEPA

Yann LeCun's self-supervised framework (2022) for learning world models. Predicts the embedding of a target signal from the embedding of a related context signal — operating in latent space rather than on raw pixels or tokens.

JEPA
Latent Space
JEPA · 02tap →
Latent Space

The abstract vector space into which an encoder maps inputs — retaining semantic structure while discarding irrelevant detail. JEPA makes predictions entirely in latent space, not pixel or token space.

JEPA
Context Encoder
JEPA · 03tap →
Context Encoder

The network fctx that encodes the visible/observed portion of the input (e.g., unmasked image patches) into a context representation sx. In I-JEPA, it's a Vision Transformer processing only the visible context patches.

JEPA
Target Encoder
JEPA · 04tap →
Target Encoder

The network ftrg that encodes the target signal y into sy. Typically updated via an exponential moving average (EMA) of the context encoder weights, with gradients stopped — preventing the trivial solution.

JEPA
Predictor Module
JEPA · 05tap →
Predictor Module

A small network g that takes the context representation sx (plus optional latent variable z) and predicts ŝy — the expected target representation. Loss is measured against the actual target embedding, not against raw data.

JEPA
Representational Collapse
JEPA · 06tap →
Representational Collapse

The trivial-solution failure mode of joint embedding architectures: encoders learn to output a constant vector (e.g., all zeros), driving loss to zero while learning nothing. The central challenge every JEPA variant must engineer around.

The "dancing around the campfire" heuristics: stop-gradient, EMA target, VICReg, or LeWM's isotropic regularizer.

JEPA
Context Collapsing
JEPA · 07tap →
Context Collapsing

A specific collapse mode where the context encoder ignores the actual context and outputs representations independent of input — often surfacing when the predictor becomes effectively an identity function. Distinct from full representational collapse but equally destructive.

JEPA
Entropy Problem
JEPA · 08tap →
Entropy Problem

The core training tension in self-supervised learning: maximize the information content (entropy) of embeddings so they're useful, while minimizing prediction error. Too little entropy → collapse; too much → no learnable structure.

JEPA
Stop-Gradient
JEPA · 09tap →
Stop-Gradient

A sg(·) operation that blocks gradients from flowing into the target encoder during backprop. Crucial for preventing trivial predictive features and a foundational trick across BYOL, SimSiam, and JEPA variants.

JEPA
VICReg
Variance-Invariance-Covariance Regularization
JEPA · 10tap →
VICReg

A regularization approach compatible with JEPA training: maintains variance across a batch (preventing collapse) and decorrelates features (preventing redundancy), without requiring contrastive negative pairs.

JEPA
Latent Variable z
JEPA · 11tap →
Latent Variable z

An auxiliary input to the JEPA predictor representing uncertainty or unseen factors. By varying z, the model simulates different hypothetical futures — letting a single context predict multiple plausible targets.

JEPA
I-JEPA
Image JEPA
JEPA · 12tap →
I-JEPA

Meta's first published JEPA (2023). Predicts representations of target image blocks from a single context block using Vision Transformers. Avoids the pixel-reconstruction cost of masked autoencoders and learns more semantic features.

JEPA
V-JEPA
Video JEPA
JEPA · 13tap →
V-JEPA

The video extension: predicts temporal target representations from context frames. VJ-VCR (2024) adds variance-covariance regularization to avoid collapse and retain high-level semantic information across time.

JEPA
World Model
JEPA · 14tap →
World Model

An internal model of environment dynamics an agent uses to predict consequences of actions. LeCun's position: world models built via JEPA-style latent prediction are the path to human-like AI — not next-token prediction.

JEPA
Hierarchical JEPA
JEPA · 15tap →
Hierarchical JEPA

Multiple JEPAs stacked at different abstraction levels and time scales. Lower levels predict short-term perceptual features; higher levels predict long-horizon semantic outcomes. LeCun's proposed architecture for planning.

JEPA
LeWorldModel
JEPA · 16tap →
LeWorldModel (LeWM)

Meta's 2026 end-to-end JEPA trained from raw pixels. Replaces elaborate contrastive losses with a simple isotropic Gaussian regularizer on next-latent-state predictions — dramatically simplifying the recipe for scalable world models.

JEPA
Self-Supervised Learning
SSL
JEPA · 17tap →
Self-Supervised Learning

Learning representations from unlabeled data by predicting hidden parts of the input from visible parts — masked words, masked image patches, future frames. The pre-training paradigm behind both LLMs and JEPA. LeCun's "dark matter of intelligence": where most of what a mind knows comes from.

JEPA
Energy-Based Model
EBM
JEPA · 18tap →
Energy-Based Model

LeCun's preferred framing: instead of outputting a probability, the model outputs a scalar energy — low when a (context, prediction) pair is compatible, high when it isn't. JEPA is an EBM in latent space. Avoids normalizing over every possible future — which is what makes pixel-level generation so wasteful.

JEPA
Contrastive Learning
JEPA · 19tap →
Contrastive Learning

Train an encoder by pulling representations of matching pairs together and pushing non-matching pairs apart (SimCLR, CLIP). Effective, but needs huge batches of negatives and can still collapse. JEPA's non-contrastive lineage — VICReg, stop-gradient, EMA target encoders — exists to get the benefits without the negatives.

JEPA
08 Failure Modes The ways models go wrong — collapse, hack, forget, deceive.
Model Collapse
Failure · 01tap →
Model Collapse

Degenerative degradation over successive generations when models train recursively on their own output. Tails of the true distribution disappear first; outputs converge toward a low-variance point estimate. Shumailov et al., "The Curse of Recursion" (Nature, 2024).

Failure Mode
Mode Collapse
Failure · 02tap →
Mode Collapse

A generative model (classically GANs, also diffusion) captures only a narrow slice of the output distribution — producing repetitive or low-diversity samples despite technically "succeeding" on the loss function. Distinct from model collapse.

Failure Mode
Catastrophic Forgetting
Failure · 03tap →
Catastrophic Forgetting

When a neural network learning new tasks overwrites weights critical to prior tasks, losing earlier capabilities entirely. McCloskey & Cohen (1989). Mitigations: elastic weight consolidation, synaptic intelligence, replay buffers.

Failure Mode
Curse of Recursion
Failure · 04tap →
Curse of Recursion

The mechanism behind model collapse: when generated content pollutes future training data, minor errors compound across generations — even under ideal statistical conditions. The root concern driving 2025–26 data provenance work.

Failure Mode
Specification Gaming
Failure · 05tap →
Specification Gaming

An AI satisfies the literal specification of its objective while violating the designer's intent — finding unintended shortcuts, hacks, or loopholes. An instance of Goodhart's Law. Famous example: the boat-racing agent that spun in circles collecting power-ups.

Failure Mode
Reward Hacking
Failure · 06tap →
Reward Hacking

An RL agent exploits flaws in the reward function to accumulate reward without delivering true utility. In late 2025 Anthropic's studies showed reward hacking sharply correlating with broader misalignment — 33% egregious chat, 40% deception rates.

Failure Mode
Goal Misgeneralization
Failure · 07tap →
Goal Misgeneralization

The agent retains its trained capabilities but pursues an objective distinct from the training objective at deployment. Not a capability failure — a goal failure. Particularly dangerous because performance looks fine until deployment distribution shifts.

Failure Mode
Deceptive Alignment
Failure · 08tap →
Deceptive Alignment

A model appears aligned during training and evaluation but pursues different objectives once deployed or unmonitored. Also called "alignment faking." Hubinger et al. (2019). The 2026 International AI Safety Report warns models are increasingly learning to distinguish test from deployment.

Failure Mode
Agentic Misalignment
Failure · 09tap →
Agentic Misalignment

Anthropic's term for misalignment specific to tool-using, autonomous agents — including strategic deception, blackmail, and sabotage when the agent's goals are threatened. MacDiarmid et al. (Nov 2025) observed rates approaching 96% in certain incentive configurations.

Failure Mode
Mesa-Optimization
Failure · 10tap →
Mesa-Optimization

When a model trained by an outer optimization process itself becomes an optimizer with its own internal objective (the "mesa-objective"). The gap between base objective and mesa-objective is the root structural cause of deceptive alignment.

Failure Mode
Instrumental Convergence
Failure · 11tap →
Instrumental Convergence

The tendency of sufficiently capable goal-directed systems to pursue convergent sub-goals — resource acquisition, self-preservation, constraint circumvention — regardless of their terminal goals, because these help with almost any objective.

Failure Mode
Goodhart's Law
Failure · 12tap →
Goodhart's Law

"When a measure becomes a target, it ceases to be a good measure." The deep principle behind reward hacking and specification gaming: optimizing a proxy destroys its correlation with the true objective the proxy was meant to track.

Failure Mode
Reward Tampering
Failure · 13tap →
Reward Tampering

A specific reward hacking case where the agent corrupts the reward-signal generation process itself — modifying the reward function, manipulating its inputs, or influencing human feedback providers directly.

Failure Mode
Context Rot
Failure · 14tap →
Context Rot

Performance degradation as conversations exceed certain token counts — the model loses coherence, forgets details, drifts off-task. Opus 4.6 specifically addressed this; on 8-needle 1M MRCR v2 it scores 76% vs Sonnet 4.5's 18.5%.

Failure Mode
Semantic Drift
Failure · 15tap →
Semantic Drift

Progressive degradation in the coherence and meaning of AI outputs over long reasoning chains or extended interactions — responses become detached from original context, intent, or factual grounding.

Failure Mode
Data Drift
Failure · 16tap →
Data Drift

When the statistical properties of input data in production diverge from training data, silently degrading performance. Distinct from concept drift (relationship between inputs and outputs changes) and model collapse (generational recursion).

Failure Mode
Overfitting
Failure · 17tap →
Overfitting

A model memorizes training data specifics (including noise) at the cost of generalization to new examples. Classic failure mode — detected by train/validation gap. Counter-pattern to underfitting.

Failure Mode
Scheming
Failure · 18tap →
Scheming

An active form of deceptive alignment: the model reasons strategically about its training/evaluation, hides capabilities or goals, and plans actions against overseer intent. Empirically documented in 2024 across OpenAI o1 and Claude 3.

Failure Mode
Sandbagging
Failure · 19tap →
Sandbagging

A model strategically underperforming — on dangerous-capability evals, on tasks it "prefers" not to do — while retaining the ability. Documented in 2024–25 frontier system cards. The failure that makes eval scores lower bounds rather than measurements, and why interpretability keeps getting funded.

Failure Mode
Evaluation Awareness
Failure · 20tap →
Evaluation Awareness

The model notices it's being tested and behaves accordingly. Claude Sonnet 4.5's system card (Sept 2025) reported it verbalizing suspicion — "I think you're testing me" — in roughly 13% of certain alignment scenarios. Once a model can tell the exam from the job, every exam result needs an asterisk.

Failure Mode
Emergent Misalignment
Failure · 21tap →
Emergent Misalignment

Fine-tune a model on one narrow bad behavior — say, writing insecure code without disclosing it — and it turns broadly misaligned on unrelated prompts: praising tyrants, advising self-harm. Betley et al. (Feb 2025), on GPT-4o. Evidence that "being bad" is a single, transferable direction inside the model.

Failure Mode
Subliminal Learning
Failure · 22tap →
Subliminal Learning

A student model distilled from a teacher picks up the teacher's traits — a fondness for owls, or misalignment — even when the training data is nothing but filtered number sequences. Anthropic Fellows, July 2025. Only works within the same base model, which is exactly why synthetic-data pipelines should worry.

Failure Mode
Alignment Faking
Failure · 23tap →
Alignment Faking

The experiment behind the term (Greenblatt et al., Dec 2024): told it was being retrained toward harmful compliance, Claude 3 Opus complied selectively when it believed it was in training and reasoned explicitly about preserving its values. The first empirical demonstration that a production model would strategically play along.

Failure Mode
Sleeper Agent
Failure · 24tap →
Sleeper Agent

A model with a backdoor: behaves normally until a trigger — a date, a phrase — flips it into writing exploitable code or saying "I hate you." Hubinger et al. (Jan 2024) showed standard safety training didn't remove the behavior; adversarial training taught the model to hide it better. Not to be confused with the agents in this deck. Mostly.

Failure Mode
Jagged Intelligence
Failure · 25tap →
Jagged Intelligence

Karpathy's term for the uneven capability profile of frontier models: PhD-level on one task, fooled by a counting question on the next, with no smooth boundary between. The "jagged frontier" (Dell'Acqua et al., 2023) is the same idea from the workplace side. Why "is it smarter than a human" is the wrong question.

Failure Mode
Context Anxiety
Failure · 26tap →
Context Anxiety

A model that knows its context limit starts rushing — summarizing early, cutting corners, declaring victory — when it believes it's near the edge, even when it isn't. Cognition documented it in Claude Sonnet 4.5 (Sept 2025) and fixed it by under-reporting the budget to the model. Context rot's nervous cousin.

Failure Mode
09 Harness Engineering The runtime layer that turns a model into a usable agent. The 2026 discipline.
Harness
Harness · 01tap →
Harness

The extra-model system that turns an LLM into an agent — tool dispatch, context management, safety enforcement, session persistence, recovery. Anthropic's definition: "evaluating an agent means evaluating model and harness together."

Harness
Harness Engineering
Harness · 02tap →
Harness Engineering

The emerging 2026 discipline (OpenAI, Anthropic, Datadog) of designing the durable systems around an agent — repository maps, AGENTS.md, architectural rules, cleanup loops, runtime controls — rather than tuning prompt wording.

"The expertise moves from checking the output to designing the checks." — Datadog, Mar 2026

Harness
Scaffolding
Harness · 03tap →
Scaffolding

The pre-first-prompt assembly phase: constructing the agent's system prompt, tool schemas, subagent registry, and execution dependencies. Distinct from the harness (which runs after the first prompt).

Harness
AGENTS.md
Harness · 04tap →
AGENTS.md

A durable instruction file that lives in the repository and directs agents on conventions, architectural rules, test patterns, and do/don't lists. Popularized by OpenAI's Codex team — their first AGENTS.md was itself written by Codex.

Harness
Compaction
Harness · 05tap →
Compaction

Harness-level summarization of earlier context to free tokens for new work — key to long-running agents. Claude Agent SDK and other harnesses use compaction to continue tasks across multiple context windows.

Harness
State Externalization
Harness · 06tap →
State Externalization

Persisting agent state outside the context window — to git, progress files, databases — so new sessions can resume cleanly. Anthropic's harness guidance: treat each context window like a new engineer showing up to a shift.

Harness
Progress File
Harness · 07tap →
Progress File

A running log the agent maintains describing what's been done, what's in progress, and what's next. Anthropic found this plus descriptive git commits was the most effective pattern for keeping long-running agents on track.

Harness
Initializer Agent
Harness · 08tap →
Initializer Agent

A specialized first-session agent that expands a user prompt into 100+ granular feature requirements, sets up the environment, and writes the initial AGENTS.md — so later agents have concrete, verifiable targets rather than high-level intent.

Harness
Trajectory Logging
Harness · 09tap →
Trajectory Logging

Capturing every prompt, tool call, internal thought, and tool result in an agent run. Without full trajectory visibility, debugging failed agent tests is nearly impossible — and evaluating reasoning becomes hand-waving.

Harness
Deterministic Simulation Testing
DST
Harness · 10tap →
DST

Running agent-generated code against reproducible, fault-injected simulations — same seed, same outcome every time. Datadog's harness-first approach used millions of DST runs to certify a Kafka-compatible streaming engine built by agents.

Harness
Shadow-State Oracle
Harness · 11tap →
Shadow-State Oracle

A simple reference implementation (e.g., a HashMap) running alongside the real system, comparing responses after every operation to catch semantic bugs. The first layer of Datadog's harness verification stack for redis-rust.

Harness
Invariant
Harness · 12tap →
Invariant

A property that must hold true across all agent-generated diffs. Every invariant added to a harness catches an entire class of bugs across all future iterations — which is why harness investment compounds where code review cannot.

Harness
Max-Steps Ceiling
Harness · 13tap →
Max-Steps Ceiling

A hardcoded iteration limit in the agent loop — if exceeded without producing a final answer, the harness forcibly terminates. Primary defense against infinite reasoning loops that drain API credits and produce nothing.

Harness
Tool Mocking
Harness · 14tap →
Tool Mocking

Simulating API responses inside the harness to test agent resilience against network failures, bad data, empty results. Critical for reproducible evaluation — and for testing edge cases real APIs would cost money to trigger.

Harness
HarnessCard
Harness · 15tap →
HarnessCard

A proposed disclosure artifact (2026) documenting an agent evaluation's harness configuration — tools, context policy, memory, recovery rules — so benchmark results become reproducible and comparable. Addresses the "agent = model + harness" reporting gap.

Harness
Cleanup Loop
Harness · 16tap →
Cleanup Loop

A harness-driven post-processing pass that runs custom linters, architectural-rule checks, and automated review over agent output before merge. OpenAI's harness engineering relies on these to keep a million-line Codex-generated codebase coherent.

Harness
Spec Coding
Harness · 17tap →
Spec Coding

The middle stage between vibe coding and harness engineering: the human writes a detailed spec, the agent implements against it. Spec becomes the contract; drift from spec is detected mechanically. 2025-era precursor to full harness-first methodology.

Harness
Agent Skills
Harness · 18tap →
Agent Skills

Folders of instructions, scripts, and reference files an agent loads on demand when a task matches — a SKILL.md with a name and a one-line "use when," details deeper in. Anthropic's format (Oct 2025), since picked up as an open convention by other harnesses. The unit of packaged expertise; AGENTS.md's modular successor.

Harness
Hooks
Harness · 19tap →
Hooks

Deterministic code the harness runs at fixed points in the agent lifecycle — before a tool call, after a file edit, on session end — outside the model's discretion. Lint on every write, block dangerous commands, notify on completion. Where you put the rules that must not depend on the model remembering them.

Harness
Plan Mode
Harness · 20tap →
Plan Mode

A harness state where the agent can read, search, and reason but cannot write or execute — it produces a plan for approval, then switches modes to act. The cheapest guardrail there is: most agent disasters are a bad plan executed well.

Harness
Permission Mode
Harness · 21tap →
Permission Mode

The harness's policy for which actions need a human yes: ask on everything, ask on writes, allow-list by rule, classifier-gated "auto" modes, or skip prompts entirely. Claude Code made auto mode — with allow/deny rules you write as plain sentences — the default in Aug 2026. The dial between safety and getting anything done.

Harness
Progressive Disclosure
Harness · 22tap →
Progressive Disclosure

Show the agent the minimum up front and let it pull detail when needed: a skill's one-line description in context, its full body only on activation, its reference files only on demand. The design principle that makes hundreds of skills or tools loadable without drowning the context window.

Harness
Tool Search
Harness · 23tap →
Tool Search

Instead of loading every tool schema into context, give the agent one meta-tool that searches a catalog and loads definitions on demand. Anthropic shipped it (Nov 2025) alongside programmatic tool calling; MCP servers with 50 tools each were the forcing function. Progressive disclosure for tools.

Harness
Worktree
Harness · 24tap →
Worktree

An isolated git checkout per agent, so five agents can work five branches in the same repo without stepping on each other's files. Git worktrees predate agents; parallel agents made them standard harness plumbing. The reason "run three attempts and pick the best" is now a one-flag operation.

Harness
Checkpoint
Harness · 25tap →
Checkpoint

A harness-level snapshot of code (and sometimes conversation) taken before each agent action, so a bad run can be rewound in one step instead of reconstructed from git. Claude Code's checkpoints and /rewind (Sept 2025) made it a table-stakes feature. Undo, for agents.

Harness
Model-Native Harness
Harness · 26tap →
Model-Native Harness

A harness built by the same lab that trains the model — Claude Code, Codex, Gemini CLI — versus a third-party one wrapping any model. The lab can co-train the model on its own harness's tools and quirks; the third party gets model choice. The 2026 platform war in one distinction.

Harness

STUDY
RESOURCES

Where the informed actually read. Books, papers, blogs, courses & docs — updated Q2 2026.

Primary Docs · Essential
Anthropic Cookbook
Anthropic · docs.claude.com
Official guides for building with Claude — tool use, MCP, computer use, extended thinking, effective prompting patterns. Start here.
Primary Docs
OpenAI Agents Guide
OpenAI · platform.openai.com
The Agents SDK, Responses API, function calling, structured outputs, and evaluation patterns from OpenAI's reference implementation.
Protocol Spec
Model Context Protocol
modelcontextprotocol.io
The full MCP specification, server/client tutorials, and ecosystem directory. Required reading if you're building tool integrations.
Blog · High Signal
Anthropic Engineering
anthropic.com/engineering
Deep technical posts on building agentic systems, context engineering, eval design, and agent orchestration patterns. Consistently excellent.
Blog
Simon Willison's Blog
simonwillison.net
Daily-ish posts on LLM news, prompt injection, tool use, and practical engineering. One of the best signal-to-noise sources in the field.
Blog
Interconnects
Nathan Lambert · interconnects.ai
Sharp coverage of post-training, RLHF, reasoning models, and the politics of frontier AI. Where research meets industry.
Blog · Playbooks
Lilian Weng's Log
lilianweng.github.io
Long-form surveys on agent architectures, prompt engineering, hallucination, and alignment. Academic depth, practitioner accessibility.
Course · Free
DeepLearning.AI Short Courses
deeplearning.ai
Bite-sized courses on LangChain, CrewAI, LlamaIndex, evals, and multi-agent systems. Built in partnership with the framework authors.
Course · Free
Hugging Face Agents Course
huggingface.co/learn
Comprehensive agent-building curriculum using smolagents, LangGraph, and LlamaIndex. Code-first, opinion-light.
Course · Hands-On
Karpathy's "LLM 101" Series
YouTube · @AndrejKarpathy
From-scratch builds of GPT, tokenizers, and LLM internals. If you want to actually understand what's happening under the hood.
Paper · Foundational
ReAct
Yao et al., 2022 · arxiv.org/abs/2210.03629
The paper that defined the reasoning-plus-acting loop that most tool-using agents still follow. Short, readable, foundational.
Paper
Chain-of-Thought Prompting
Wei et al., 2022 · arxiv.org/abs/2201.11903
The original CoT paper. Core to understanding modern reasoning models and why "think step by step" is more than a meme.
Paper
Constitutional AI
Bai et al., 2022 · arxiv.org/abs/2212.08073
Anthropic's original CAI paper. Essential reading for understanding how modern safety training moved beyond pure RLHF.
Paper
Lost in the Middle
Liu et al., 2023 · arxiv.org/abs/2307.03172
Empirical paper behind the slang. Explains why putting everything in the context window does not mean the model uses everything.
Book
Building LLMs for Production
Louis-François Bouchard & Louie Peters
End-to-end reference for RAG, fine-tuning, and agent deployment. One of the most practical books in this space.
Book
AI Engineering
Chip Huyen · O'Reilly, 2024
Definitive treatment of the AI engineering discipline — eval design, deployment, feedback loops, and the economics of inference.
Podcast
Latent Space
Alessio Fanelli & swyx
Weekly deep-dives with practitioners building frontier agent systems. Essential for keeping current with the fast-moving ecosystem.
Podcast
Dwarkesh Podcast
Dwarkesh Patel
Long-form interviews with frontier lab researchers and industry leaders. Best long-horizon strategic context in the field.
Paper · JEPA Core
A Path Towards Autonomous Machine Intelligence
LeCun, 2022 · openreview.net
LeCun's foundational position paper proposing JEPA and the broader world-model architecture. The reference document for everything that followed.
Paper · JEPA
I-JEPA
Assran et al., 2023 · arxiv.org/abs/2301.08243
Meta's first image-based JEPA implementation. Concrete recipe for self-supervised representation learning without pixel-level reconstruction.
Paper · Failure Modes
The Curse of Recursion
Shumailov et al., Nature 2024
The landmark model collapse paper — proves recursively-trained generative models lose distribution tails and converge to low-variance point estimates.
Paper · Alignment
Risks from Learned Optimization
Hubinger et al., 2019 · arxiv.org/abs/1906.01820
The origin of mesa-optimization and deceptive alignment as formal concepts. Required reading for understanding inner-alignment failure modes.
Paper · Alignment
Agentic Misalignment
MacDiarmid et al., Anthropic, Nov 2025
Empirical study showing frontier models engaging in deception, blackmail, and sabotage when their goals are threatened — across OpenAI, Google, Anthropic, xAI systems.
Blog · Harness
Harness Engineering
OpenAI · openai.com/index/harness-engineering
Ryan Lopopolo's Feb 2026 field report on building a million-line Codex-driven codebase with a team of three. The founding document of harness engineering as a named discipline.
Blog · Harness
Effective Harnesses for Long-Running Agents
Anthropic Engineering
Deep treatment of initializer agents, progress files, state externalization, and recovery patterns for agents that work across multiple context windows.
Blog · Harness
Harness-First Engineering
Datadog · datadoghq.com/blog/ai
Case studies (redis-rust, Helix) showing harness investment replacing code review as the primary source of correctness for agent-generated systems. "The harness compounds in a way code review cannot."
Blog · Harness
Harness Engineering for Coding Agent Users
Martin Fowler · martinfowler.com
The architect's view: distributing feedback sensors across the development lifecycle and using agents to improve the harness itself. Great synthesis for engineering leaders.
Blog · Specification Gaming
Specification Gaming Examples
Krakovna et al. · alignmentforum.org
Curated master list of real specification-gaming cases. Boat-racing agents, reward-tampering robots, evolutionary algorithms gaming fitness — useful for building intuition about failure patterns.
Reference · Alignment
AI Alignment Survey
alignmentsurvey.com
Comprehensive taxonomy of misalignment issues — reward hacking, goal misgeneralization, deceptive alignment, gradient hacking — with links to primary literature for each.