Agent bugfix

Load a Pi coding agent into a sandbox, hand it a failing test, and stream every tool call as it debugs and fixes it.

alineo runs Pi coding agents inside sandboxes. You hand the agent a task and stream back everything it does — text, and every tool call — while it reads files, edits code, and runs commands inside its own container.

This workflow plants a buggy fib.py with a failing test, hands it to a python-data agent, and streams the session until the agent stops. Then it runs the test suite independently to check the agent's work.

Not connected

This workflow loads a Pi agent, which needs a model API key on the backend. The hosted endpoint has one configured; a local server needs GEMINI_API_KEY in its environment.

createAgentprompt (stream)deleteAgent~2 min
Connect to a sandbox server above to run this.

What just happened

  • Alineo.load(pythonDataSpec) booted an agent — a sandbox with Pi installed and configured.
  • agent.prompt(task) returned an event stream. Each tool_start / tool_end is the agent running a real command or file edit inside its sandbox; text is the model's narration.
  • The verify step ran pytest ourselves — never take the agent's "all green" on faith.
  • agent.close() ended the Pi session and deleted the sandbox.

The AI Agent Bugfix cookbook adds an independent verifier agent and a structured verdict. For streaming mechanics, see Streaming.

Fork a sandbox

Do the expensive setup once, then fork the live container into independent children.

On this page

What just happened