Use fastaiagent with a coding assistant¶
A coding assistant — Claude Code, GitHub Copilot, Cursor, Google Antigravity and the
rest — was not trained on this SDK's current API. Left alone it guesses: it calls
connect() without a target, imports the LangChain adapter that was removed in 1.96.0,
or reaches for discover_tools(). The code looks right and fails when you run it.
The SDK ships two Agent Skills. A skill is a folder with a SKILL.md that the assistant
reads when the work calls for it.
fastaiagent-sdkis read when the assistant is about to writefastaiagentcode. It holds the ten mistakes assistants make most, a reference per area (agents, tools, teams, guardrails, evals, platform, testing, what changed), a script that checks what the assistant wrote, and a short recipe for starting a project: a layout where plainpytestworks and an entry point that connects and always ends withdisconnect().
It also asks the assistant to do more than write code that passes its offline tests. Those tests
use scripted models, so they cannot see a wrong model id, a missing key or a dead server. When
your model key is in the environment the assistant should run the real entry point once and say
what it ran and what it did not. It should keep an address you give it in FASTAIAGENT_TARGET
instead of in code, test an MCP tool against a real local server instead of a fake, and not add
tools or helpers you did not ask for.
For an app connected to a platform, the skill says what the platform supplies (guardrail rules,
prompts, eval datasets, judges, knowledge-base search) and what it receives (traces, eval
verdicts). It also gives a checklist the assistant runs before calling the app done: that the
plane's guardrail rules reached the agent, that one of them actually blocks, that the trace came
back from the platform, and that a gated eval run was sent.
- fastaiagent-studio is read when you ask for an agent to be built through
Agent Studio. It teaches the fastaiagent studio commands and how
to write a spec (see The Studio skill below). It needs a platform.
Install it¶
That copies both skills into the two folders assistants read:
| Folder | Read by |
|---|---|
.claude/skills/ |
Claude Code. GitHub Copilot and Cursor read it too. |
.agents/skills/ |
GitHub Copilot, Cursor, Google Antigravity, Codex, Gemini CLI and others |
The command prints what it did as JSON and exits non-zero if it could not do what you
asked. It also offers to add a three-line pointer to AGENTS.md (created if missing,
appended to otherwise, never rewritten); it asks only when it has a terminal. Pass
--agents-md or --no-agents-md to decide up front, and --claude-md to add the pointer
to CLAUDE.md too — it never edits that file otherwise.
| Command | Does |
|---|---|
fastaiagent skills list |
The packaged skills, and any installed copies that are out of date |
fastaiagent skills install |
Install into the project. --tool claude\|agents\|all, --user for every project, --force to replace |
fastaiagent skills path |
The folder the skills ship in |
Each installed copy is stamped with the SDK version that installed it. After you upgrade the
SDK, skills list says which copies are stale and skills install --force refreshes them.
Install it without the SDK¶
The same skills are in a public repository,
fastaifoundry/fastaiagent-skills:
Copy a folder from skills/ (fastaiagent-sdk/, fastaiagent-studio/) into .claude/skills/
or .agents/skills/ in your project, or into ~/.claude/skills/ or ~/.agents/skills/ for
every project.
The repository is generated from the SDK at each release; it does not take pull requests. The
gh skill install line above names one skill; use fastaiagent-studio for the other.
What the assistant does with it¶
- You ask for
fastaiagentcode. - The assistant loads
SKILL.mdand reads the reference for the area — for examplereferences/tools.mdbefore it writes a tool. - It writes the code and an offline test using
fastaiagent.testing.TestModel. - It runs the checker from the skill's folder:
The checker resolves every fastaiagent import and every fa.something attribute against
the installed SDK, and names the replacement for anything that was removed:
agent.py:3: error: cannot import discover_tools from fastaiagent.tool (removed. Use MCPTool.from_server(...) (needs the [mcp] extra).)
agent.py:7: warn: connect() needs target=... (or FASTAIAGENT_TARGET in the environment); there is no default since 1.100.0.
checked 1 file(s): 1 error(s), 1 warning(s)
It exits 0 when clean, 1 on an error (or on a warning with --strict), and 2 if the SDK
is not importable from that Python. You can run it yourself, or in CI.
The Studio skill¶
Agent Studio can build an agent for you from a spec. With the
fastaiagent-studio skill installed, ask your assistant for an agent ("build me an order-status agent
with Studio, here is the orders data") and it follows this loop:
fastaiagent studio context— which LLM endpoints and knowledge bases the project has. With no endpoint it stops and tells you to add one in the console.- It reads what you want, and any requirements document or scan itself, and says what it could not read.
- It lays the design out for you — the shape, each worker, the tools, the guardrails, the eval cases — and waits for a yes. The API has no plan step, so this is where you approve; nothing is pushed before it.
studio newwith your sample data, then it writes the spec and runsstudio spec validate. Every warning is read out to you.studio spec push --wait, thenstudio show,studio tryandstudio eval.- To change something it reads the spec back, edits it and pushes a new version; the old one is kept.
studio exportinto your repository, then it runs the exported tests and tells you how to run the agent with your own key.
It never prints or stores a key, reports a refusal and its reason instead of retrying it, tries a team one worker at a time, and stops and asks if someone changed the draft in the console. It needs a platform at 0.20.0 or later and an API key with the Agent Studio permissions.
This skill has not been tried by hand in GitHub Copilot, Cursor or Google Antigravity yet, so this page does not claim it works there: it is in the folder each of them reads, and the checklist for trying it is in the SDK repository.
What it does not do¶
- It does not make an assistant reliable on its own. It removes one cause of wrong code — a guessed API — and gives you a check for it. Read what the assistant writes.
- It does not call the model, the platform or the network. The checker runs none of your code.