Skip to content
Chapter 27 of 28
Illustrated overview of Observe and operate agents. An accessible explanation follows below.
Enlarge / view original infographic The key ideas are also explained in the text below.

Observe and operate agents

Combine evidence, bounded retries, safe tool access, secrets management and escalation.

Keep the minimum useful evidence

The infographic lists a structured plan, commit and pull-request history, workflow results, artifacts and review outcomes. Together these help explain what happened and why a result was accepted or rejected.

Preserve necessary logs and reports using a controlled retention policy. Record decisions and metadata rather than indiscriminately logging prompts, credentials or sensitive tool output.

Design for imperfect execution

Bound retries and distinguish transient failures from invalid requests or missing authorization. Repeated failure should trigger escalation, not an endless loop. Prepare a rollback approach for high-impact changes.

Keep capability narrow

Explicit tool allowlists, restricted MCP access and minimal workflow permissions reduce the blast radius. Secrets belong in protected mechanisms, not instructions or committed files. Pass only what is needed to the relevant environment.

The hook JSON and permission fragments in the image are conceptual; their event schema and runtime behavior are not a universal implementation. Use the current hooks reference and test fail-closed behavior before relying on a control.