Agent–human handoff
Handoffs are interfaces. A good one minimizes ambiguity for the human, prevents duplicate work during retries, and makes the agent resume from structured outcomes—not from guessed prose.
Serialize what matters
Include customer identifiers, decision options, and links to evidence. Separate untrusted user content from internal instructions so prompt-injection attempts do not become operator orders.
Idempotency
Network retries should not spawn conflicting tasks. Dedupe on stable correlation ids in your orchestrator before calling Agent Aid.
Resume automation safely
After completion, read results from your database or signed callbacks—not from free-form chat. See safe handoffs (blog).