Receipts-First Agent Launch Checklist (for 2026 buyers)

Goal: Ship an agent that completes a scoped job end-to-end and leaves behind an audit trail that an operator can use to debug, approve, and rollback.

1) Define the job boundary (non-negotiable)
- Name the job in one line: verb + object + system (e.g., “Reconcile vendor invoices into NetSuite”).
- List allowed inputs and outputs. Anything not on the list is out of scope.
- Write the top 5 failure modes you expect (missing permissions, ambiguous data, tool outage, rate limits, conflicting human edits).

2) Identity and permissions
- Decide: per-user identity (OAuth/SSO) vs service account. Default to per-user.
- Document required scopes for each integration.
- Implement least privilege: separate read scopes from write scopes.
- Add an “impersonation” story for admins (with explicit logging) if your buyer needs it.

3) Receipts (audit log) design
- Use an append-only event log per run: planning events, tool-call events, approval events, and finalization events.
- Store: tool name, action, inputs hash, idempotency key, outputs summary, timestamps, actor identity, and policy decision.
- Make logs searchable in-product by run_id, user, tool, and external object ID.

4) Safety controls
- Add approval gates for: money movement, external messages, privilege changes, and production writes.
- Provide operator actions: pause, resume, edit step, skip step, rerun step.
- Implement rollback where the target system supports it; otherwise, implement compensating actions and document limitations.

5) Reliability engineering
- Make side-effecting calls idempotent (idempotency keys, de-dupe checks).
- Standardize retries with backoff; classify errors (auth, quota, validation, transient).
- Add a dead-letter queue or “needs human” state for stuck runs.

6) Evaluation and regression
- Build a fixed scenario suite: real-ish examples with expected tool calls and expected outputs.
- Run evals on every prompt/tool change.
- Save “golden traces” (the sequence of steps) so you can spot regressions in behavior, not just final text.

7) Rollout plan
- Start read-only in production systems; graduate to write access behind approvals.
- Ship feature flags per customer, per integration, and per action type.
- Provide an admin dashboard: permissions, logs, approval policies, and usage meters.

Definition of done: An operator can answer (1) what the agent changed, (2) why it changed it, (3) who approved it, and (4) how to undo it—using your product alone.