Agent control plane with guarded actions

Give every action an owner, a boundary and a way to stop.

Once an Agent can inspect jobs, change configuration or trigger a rerun, accuracy is only one question. The platform also needs to know which user authorised the action, which tools were available, what was written and how to revoke the path.

We delivered the control surface as replaceable runtime ports: OAuth identity, scoped tool policy, Guardrails, conversation isolation, approval gates and end-to-end trace IDs. Capability and governance ship together instead of waiting for a security review after the pilot.

In the work

Identity before capability
Users and Agents operate through an attributable identity domain instead of a shared operations credential.
Reads wide, writes narrow
Inspection can compose multiple tools; mutations require explicit scope and approval before persistence.
Trace and revoke
Trace IDs connect reasoning and tool calls, while policy and adapter boundaries provide a practical stop point.

The problem

A multi-system data platform needed Agents to inspect and recommend changes without turning service credentials into an unowned automation account. Identity, write policy and traceability had to be part of the runtime, not a document after launch.

Measured result

The governed runtime supports a daily user group of about 60 engineers; 30%+ of task-failure questions close without on-call escalation, platform on-call volume is down 60%+, and typical SLA-chain recovery is about one hour shorter.

Scope