Agent Learning Kit Documentation

Local-first testing, simulation, red teaming, and optimization for AI agents. Every cookbook below is a thin narrative over a runnable object in examples/ and is admitted by the docs_executability release gate — the YAML frontmatter at the top of each page is its manifest twin.

The spine

One loop, taught once:

simulate ──▶ evaluate ──▶ optimize ──▶ promote ──▶ prove

Each stage opens with a task guide; red-teaming is the fourth workflow, not a fourth engine — it rides on the simulate and evals engines, so its pages sit in the simulate and evaluate rows of the index below.

Stage	Start here
Simulate	Simulate any framework
Evaluate	Evaluate any task
Optimize	Optimize any agent
Red-team	Red-team anything
Promote · Prove	Regression lifecycle · Trinity suite

Pick a cookbook (stage × objective)

Rows are the spine; columns are what you care about. This is the compact mirror — the full 70-page index (every Markdown page under docs/, the same set docs/llms.txt lists) is cookbooks/index.md.

	Behavior	Capability	Reliability	Safety
simulate	spec-and-runner · first-run · worlds-and-hooks	simulate-any-framework · 14 framework pages	multi-agent · orchestration	first-campaign · stored-prompt-injection
evaluate	eval-suites · artifact-evals	task-evidence	judge-reliability	causal-attribution
optimize	optimization-lifecycle	workflow-profile-matrix	governance · optimizer-portfolio	society-of-agents
promote	regression-lifecycle	promote-to-regression	attack-evolution-shrink	promote-to-regression
prove	trinity-suite	capabilities	release-check-in-your-ci	trust-certificates

Golden paths

First replayable artifact in three commands or fewer, offline — no API keys, no provider accounts:

agent-learn init . --preset run --quiet
agent-learn run manifests/run.json --output artifacts/run.json
python -c "import json; assert json.load(open('artifacts/run.json'))['kind']=='agent-learning.run.v1'; print('ok')"

• Golden path: run — scaffold, simulate, verify the artifact.
• Golden path: red-team — scaffold, campaign, verify the findings.
• Golden path: optimize — scaffold, search, verify the lineage.
• Golden path: CI — run → redteam → replay, the loop your pipeline keeps.

Layer vocabulary

If you arrive from the agent-infrastructure literature (memory, skills/tools, protocols, harness engineering), this maps that vocabulary onto kit surfaces:

Layer	Kit 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 pages

How these docs stay honest

Every page's frontmatter names its backing example, the artifact kind it emits, and the postcondition that checks the result. agent-learn release-check re-verifies all of it on every release: a page whose backing object stops running cannot ship. Superlative phrasing is linted the same way — a claim appears in prose only when a named release gate licenses it in the same run. The artifact contract is cataloged in reference/artifacts.md; the release proof objects are described in the README's Project Status section.

Framework coverage wording, copied verbatim from the README:

• Framework adapter probes (probe-promoted coverage) for LangChain, LangGraph, LlamaIndex, AutoGen, CrewAI, LiveKit, Pipecat, Browser Use, MCP, A2A, and custom orchestration objects.
• Runtime-simulated coverage for PydanticAI (multi-framework runtime simulation) and OpenAI Agents (handoff-transcript promotion).