Everything the demo skipped

Plan, call, observe, repeat — with a checkpoint written after every step. Long runs survive deploys, restarts, and the tool that goes down for twenty minutes.

• Checkpoint after every step
• Retries with backoff and jitter
• Resume days later from the same state
• Cancel and rewind to any step

run.resume('8f21c') # picks up at step 3, not step 0

Write an ordinary function. Pilot reads its signature, derives the JSON schema, and validates every call before your code runs — so a hallucinated argument is a caught error, not a production incident.

• Schema derived from the signature
• Arguments validated before execution
• Enums and value objects supported
• Tools versioned alongside your code

#[Tool] function refund(string $chargeId, int $amountCents, Reason $reason): Refund

Working memory, scratchpads, and intermediate results live in your database, not in a variable. Nothing is lost to a restart, and two workers never disagree about where a run got to.

• Postgres, SQLite, or your own driver
• Optimistic locking across workers
• Scratchpads scoped per run
• Automatic pruning on completion

Some decisions should not be automatic. Drop an approval gate into the loop and the run parks itself, routes to the right person, and continues the moment they sign off.

• Gates anywhere in the loop
• Routing to Slack, email, or your own UI
• Runs can wait for days
• Every approval recorded in the trace

await approval('refunds')->from($team)->because('over $500')

Every prompt, argument, result, token, and cent is recorded per step. Replay a run against a different model or a changed prompt and diff the two side by side before you ship the change.

• Full step-level recording
• Replay against a new model or prompt
• Diff two runs step by step
• OpenTelemetry export

Promote any recorded run to a test case. Your suite runs on every pull request and tells you which step regressed, not just that the score moved.

• Recorded runs become fixtures
• Assertions per step, not per run
• Runs in GitHub Actions
• Regression attributed to a step

Cost is attributed to the step that caused it, so you find the expensive loop instead of guessing. Set a ceiling and a run halts rather than quietly billing you for an afternoon.

• Per-run and per-step attribution
• Hard budget ceilings that halt
• Cached results reused on retry
• Per-step model routing