The outcome
Publish an agent that answers or acts within a documented boundary and fails safely when knowledge, authorization, tools, or models are uncertain.
Step by step
A workflow you can repeat.
- 01
Define users, purpose, prohibited uses, data classes, knowledge owners, models, tools, channels, output and citation rules, side effects, budget, retention, escalation, and rollback.
- 02
Create an isolated project, ingest minimal approved knowledge with source and expiry metadata, choose reviewed models and plugins, and keep prompts and variables free of secrets.
- 03
Build explicit workflow inputs, branches, schemas, timeouts, retries and stopping conditions; restrict tool destinations and permissions, reauthorize resources, and require approval before consequential actions.
- 04
Create evaluations for normal, ambiguous, stale, conflicting, no-answer, multilingual, unsafe, injection, knowledge-poisoning, permission, plugin-failure, duplicate-action and token-exhaustion cases.
- 05
Gate publication on results, publish privately to a limited audience, monitor traces, feedback, cost and tool calls, refresh reviewed knowledge, revoke compromised integrations, and retain a prior version for rollback.
Working standard
What good use looks like.
- Use evaluations before publication.
- Require evidence or abstention.
- Approve consequential tool calls.
Official references