52 pages

concepts

Ideas, theories, methods, frameworks, terminology. agent-computer-interface ACI: design tools for Agent goals, not API endpoints; when-to-use descriptions; structured errors. agent-evaluation Runtime Agent eval: Pass@k vs Pass^k, grader types; eval as the RSI bottleneck. agent-legibility Designing systems for the agent's eye, not the human's; logs, metrics, code all agent-readable. agent-loop The ~20-line ReAct loop; Workflow vs Agent; five control patterns; loop as optimization target. agent-memory Four memory types (working/procedural/episodic/semantic); MEMORY.md + consolidation. agent-sandboxing Sandbox as OS-enforced execution boundary; per-OS primitives; Codex-on-Windows design axes. agentic-context-engineering ACE: context as evolving playbook; Generator/Reflector/Curator; incremental bullets only. architectural-invariants Strict layering enforced by custom linters with embedded fix instructions. chain-of-thought CoT "think step by step" + Tree of Thoughts; test-time compute; ancestor of reasoning models. chinchilla-scaling Compute-optimal data/params rule; models over-train past it; FLOPs, not params, predict quality. claude-hooks Lifecycle shell hooks; deterministic enforcement of things you don't trust the model with. claude-md The project-root contract file; short, hard, executable — not a wiki. claude-skills On-demand knowledge/workflow packages; descriptors resident, bodies progressively disclosed. claude-subagents Spawned Claude instances with isolated context; about isolation, not parallelism. codebase-as-system-of-record "If the agent can't find it in the repo, it doesn't exist"; versioned artifacts as truth. constitutional-ai Anthropic alignment: written constitution + AI critique + RLAIF instead of per-example human labels. context-engineering Choosing what enters the LLM context; steering-not-teaching; tiers, Context Rot, compaction. data-engineering Data recipe = capability design; "models must get bigger before they can get smaller". deliberative-alignment OpenAI alignment: model reasons about safety policy at inference time; reasoning-model-enabled. distillation Teacher-to-student capability transfer; diffusion staircase; release ≠ rightmost checkpoint. entropy-and-garbage-collection Golden principles + scheduled cleanup tasks; debt paid down in small installments. eval-grader-reward Training-time eval/grader/reward loop; ORM vs PRM; the grader is the critical failure point. evolutionary-search-llm Evolutionary optimization for LLM programs: AlphaEvolve, DGM, Promptbreeder, ADAS, AFlow. grpo Group Relative Policy Optimization; drops PPO's value network; default for verifiable-reward RL. harness Acceptance + boundary + signal + fallback around the loop; matters more than model choice. llm-agent Weng's 2023 anatomy: LLM brain + Planning + Memory + Tool use; ancestor of the wiki's agent pages. llm-training-pipeline Six-layer / nine-stage frame; the back half decides perceived capability. long-running-agents Initializer + Coding Agent split; state externalized to files; release autonomy in order. loop-engineering Anthropic's four-type loop taxonomy (turn/goal/time/proactive); the hand-off ladder. maximum-inner-product-search MIPS + ANN (LSH/ANNOY/HNSW/FAISS/ScaNN): the vector-retrieval substrate for long-term memory. meta-context-engineering MCE: bi-level search over context *mechanism* + content; free-form skills, agentic crossover. meta-harness Outer-loop optimizer that rewrites harness *programs*; same-model 6× spread. mixture-of-experts MoE: scale params without scaling per-token compute; routing and infra complexity. model-and-effort-selection Claude Code's two dials: model = capability, effort = thoroughness; triage + token economics. model-context-protocol MCP: how external systems plug into Claude Code; largest hidden context cost. multi-agent-orchestration Orchestrator + workers; JSONL inbox protocol, worktree isolation, hallucination amplification. plan-mode Read-only exploration mode; refines a plan before any file is touched. post-training SFT/RLHF/DPO/RFT routes; DeepSeek-R1 four-stage recipe; SFT teaches style as much as knowledge. pretraining Floor not ceiling; tokenizer/context/multimodal commitments lock in here; pretraining = RL priors. prompt-caching Anthropic's prefix-cache mechanism; the architectural backbone Claude Code is shaped around. prompt-injection Source-sink decomposition; tag untrusted content; independent LLM verifier; sandbox as sink-cutting. ralph-wiggum-loop Agent self-review pattern: loop until every reviewer + verifier is satisfied; Codex's PR cycle. react Reasoning + Acting; the bridge that lets language priors generalize across RL environments. reasoning-models o1 / R1 paradigm; the second scaling axis (inference compute); effort as the user-facing dial. recursive-self-improvement RSI: AI improving its own machinery; the near-term path is harness optimization, not weights. reward-hacking Reward overfitting → hacking → tampering → alignment faking; acute in self-improvement loops. self-improving-harness An LLM optimizing its own harness: STOP, Self-Harness, DGM; propose-evaluate-accept loops. self-reflection Reflexion / Chain of Hindsight / Algorithm Distillation; ancestor of verifier + ralph-wiggum loops. six-layer-agent-architecture Tw93's Claude Code frame: CLAUDE.md / Tools+MCP / Skills / Hooks / Subagents / Verifiers. task-decomposition Planning by subgoals: CoT, Tree of Thoughts, LLM+P; the decomposition half of agent Planning. tool-use MRKL/Toolformer/HuggingGPT/API-Bank; "when & how to call" is the crux; ancestor of ACI + MCP. verifier-loop The closing layer; "Claude said done" isn't done — bind acceptance criteria up front.
esc