Agentic systems
Agentic systems need recovery design, not just better prompts
Why permissions, canaries, storm guards, rollback and decision-ready escalation belong in the product design of operational AI agents.
A successful path proves too little
A demonstration shows that an agent can complete a task when the environment cooperates. Operations asks different questions: what happens when credentials expire, a dependency starts slowly, retrieved evidence is stale or the proposed action is unsafe?
Prompt quality matters, but it cannot replace service checks, permission boundaries, action caps and recovery behaviour.
Separate detection from authority
The component that notices a failure should not automatically receive unlimited power to fix it. Detection, diagnosis, recommendation, approval and execution can be separate stages with different credentials and evidence requirements.
That design sounds slower until the first false positive. In practice it permits faster automation because the allowed actions are narrow, reversible and easy to audit.
Make failure decision-ready
Good automation either restores service within a bounded policy or produces a useful escalation: what failed, what was tested, what changed, what did not change, and what decision is now required.
Canaries, rollback, restart caps and last-known-good state turn incidents into controlled transitions. They also make an agent easier to trust because its limits are visible before something goes wrong.