Architecture / 8 min read
A practical control model for agentic quality engineering
Separate planning, domain reasoning, tool execution, and approval so autonomy remains observable and accountable.
Agentic quality workflows become governable when responsibilities are explicit. The orchestrator should plan and coordinate. Domain agents should reason within declared capability contracts. Tools should perform bounded external operations. Policy should decide when an action is allowed, denied, or paused for approval.
This separation creates evidence at every boundary: the input scope, selected skill, model route, retrieved context, generated output, validation result, tool request, approval decision, and final artifact.
What to record
- Tenant, project, actor, role, environment, run, and correlation identifiers.
- Plan version, step state, selected capability, and policy outcome.
- Source provenance, retrieval sufficiency, and unresolved assumptions.
- Provider, model, tokens, latency, cost, and structured-output validation.
- External target, idempotency key, approval, result, and artifact references.
The objective is not to slow automation down. It is to make fast automation safe to understand, interrupt, recover, and improve.