For AI coding agents¶
One file teaches an AI coding agent (Claude Code, Cursor, Copilot) the SDK: what to install, which function does each job, the safety rules and the common errors. It is generated from the code, so it matches the version you installed.
flowchart LR
C["the code<br/>__all__, signatures, pyproject"] --> G["scripts/gen_agent_guide.py"]
G --> R["concise: AGENTS.md<br/>(repo root, wheel)"]
G --> L["full: llms-full.txt<br/>(docs site, wheel)"]
R --> CLI["agent-adapter guide"]
L --> F["agent-adapter guide --full"]
CLI --> P["your project:<br/>docs/agent-adapter.md"]
Two sizes¶
| Tier | Size | Has | Get it |
|---|---|---|---|
| Concise | about 26 KB (6.5k tokens) | install, import rules, safety rules, recipes, error fixes, CLI | agent-adapter guide |
| Full | about 125 KB (30k tokens) | the concise guide plus every public signature | agent-adapter guide --full, or llms-full.txt on this site |
A file your agent loads on every session costs its size in tokens each time, so give it the concise tier and let it fetch the full one when it needs a signature.
Give it to your agent¶
Write the guide to its own file, then point your agent's instruction file at it. --out
never replaces a file that exists (your own AGENTS.md stays safe) unless you add --force.
| Agent | Then |
|---|---|
| Claude Code | add the line @docs/agent-adapter.md to your CLAUDE.md |
Cursor, Codex and other AGENTS.md readers |
add "Before using Agent Adapter, read docs/agent-adapter.md." to AGENTS.md |
| GitHub Copilot | add the same line to .github/copilot-instructions.md |
| Any chat | paste the output of agent-adapter guide, or give it the full-guide link |
| A short paste | agent-adapter guide --section safety (or install, recipes, errors, cli, api) |
No AGENTS.md yet? agent-adapter guide --out AGENTS.md writes one. No script on the
PATH? python -m agent_adapter guide does the same. Output is UTF-8 on every platform,
so agent-adapter guide > file.md works in any Windows console. The docs index for
agents that browse is llms.txt at the site root; its header
links the full guide.
What is in it¶
Sections come in this order, so an agent that reads only the top still gets the rules.
| Section | What the agent learns |
|---|---|
| What it is | a tool SDK that works with no agent framework and can be your agent loop |
| Install | one pip install line per need, each extra's supported versions |
| Import rules | where each name is imported from; never agent_adapter.core |
| Safety rules | approvals only through the executor, confirm=True for side effects, no secrets in logs, no global cache for per-user data, no retries on side effects |
| I want to… → use… | about 40 short recipes: tools, secrets, sync and async, structured output, web apps, queues, approvals, budgets, cache, tracing, fan-out, flows, hosted and local models, compression, imports, MCP, HTTP, A2A, the optimiser |
| Common errors → fixes | the error catalogue; the full tier adds the Troubleshooting tables |
| CLI commands | each command as one line |
| API index (full only) | every public name with its signature and first docstring sentence, by module |
Keep it current¶
- After an upgrade, run
agent-adapter guide --out docs/agent-adapter.md --force. The packaged copy always matches the installed version. - In this repository,
python scripts/gen_agent_guide.pyrewrites all four copies.tests/docs/test_agent_guide.pyfails when one is stale or a public name is missing;tests/docs/test_agent_guide_recipes.pyfails when a recipe stops running.
Files¶
| File | Does |
|---|---|
scripts/gen_agent_guide.py |
writes the four copies (--check only compares) |
scripts/agent_guide/ |
one module per section: install, recipes, API index, prose |
agent_adapter/cli/guide.py |
agent-adapter guide: prints, writes or slices the packaged copies |
tests/docs/test_agent_guide.py |
drift, order, size, completeness, domain words, quoting |
tests/docs/test_agent_guide_recipes.py |
runs every recipe, or compiles and API-checks it |
tests/cli/test_guide.py |
the guide command and the --help examples |