Cookbook Index

Pick a row (what stage you're at) or a column (what you care about). Every cell links real pages; empty cells render as —, never invented pages. A page may appear in at most two cells.

The matrix (stage × objective)

	Behavior	Capability	Reliability	Safety
simulate	golden-path-run · spec-and-runner · first-run · worlds-and-hooks	simulate-any-framework · memory · multimodal-image · voice-realtime · langchain · langgraph · llamaindex · autogen · crewai · openai-agents · pydantic-ai · livekit · pipecat · browser-use · mcp · a2a · custom · openenv · persona-scenario-studio	multi-agent · orchestration	golden-path-redteam · red-team-anything · corpus · first-campaign · stored-prompt-injection · long-horizon · autonomous-task-world · persona-conditioned-attacks
evaluate	evaluate-any-task · eval-suites · artifact-evals	task-evidence · eval-hooks · benchmark-overview · benchmark-coding · benchmark-voice · benchmark-pull-rl · benchmark-write-a-suite	judge-reliability · persona-fidelity · benchmark-command-graded	causal-attribution · benchmark-sandboxes
optimize	golden-path-optimize · optimize-any-agent · optimization-lifecycle · behavior-and-collaboration	world-model · memory-targets · workflow-profile-matrix · optimizer-profile-matrix · agent-control-plane	governance · optimizer-portfolio · backend-routing · multi-agent-targets · eval-suite-optimization	society-of-agents · campaign-optimization · adaptive-loop
promote	regression-lifecycle	promote-to-regression	golden-path-ci · attack-evolution-shrink · capability-profile-freezing	promote-to-regression
prove	trinity-suite	actions · capabilities	release-check-in-your-ci · observability · benchmark-in-ci	trust-certificates

Notes on placement: red-team pages live in the simulate and evaluate rows because red-teaming rides the simulate and evals engines; the promote row holds the pages whose output is a regression baseline you keep (promote-to-regression appears under both capability and safety — same page, two reasons to need it).

By layer (literature vocabulary → kit surface)

Layer	Pages
memory	optimize/memory-targets · simulate/memory
skills / tools	prove/actions · redteam/first-campaign
protocol	frameworks/mcp · frameworks/a2a · frameworks/openenv
harness	prove/release-check-in-your-ci · quickstart/golden-path-ci

By framework

LangChain · LangGraph · LlamaIndex · AutoGen · CrewAI · OpenAI Agents · PydanticAI · LiveKit · Pipecat · Browser Use · MCP · A2A · custom · OpenEnv (compatibility input)

Indexes and reference

These pages carry backing: [] — they catalog the corpus rather than run it, so they sit outside the matrix:

• Landing page — the spine, golden paths, and the compact mirror of this matrix.
• This page — the full 2-axis index.
• reference/artifacts.md — the closed artifact-kind catalog.
• reference/cli.md — the agent-learn command surface.
• reference/configure.md — API key semantics, offline vs platform.

How to read a page

Every cookbook page follows one skeleton: frontmatter (the manifest twin), what you are testing, run it (CLI and SDK for the same operation), what you built (the machine-checkable postcondition), when it fails (symptom → layer → doctor check), and the next spine step. The frontmatter declares the backing example under examples/ and the artifact kind the page emits; the docs_executability release gate re-verifies both on every release, so a page whose runnable twin breaks cannot ship. Pages flagged opt_in_lane: true (voice and live-provider sessions) are not on the golden path and link the follow-up phase that covers their live infrastructure.