Responsibility & ownership
Explicit human accountability for agents, tasks, the platform, and shared capabilities.
Four roles, clear accountability
Every Agent, Task, and Capability needs a named human or accountable team. These roles may be held by the same person in a small team, but the responsibilities remain distinct.
| Role | Accountability | Example failure |
|---|---|---|
| Agent Owner | Agent purpose, guardrails, behavior, quality across sessions (not a particular session outcome), and required operating permissions. | Agent repeatedly produces incorrect analyses |
| Task Owner | Task objective, delegation grant, acceptance criteria, and business outcome. | Unauthorized task or unacceptable result |
| Platform Owner | Harness, execution infrastructure, observability, and policy enforcement. | Unauthorized tool call is not blocked |
| Capability Owner | Correctness, reliability, and access restrictions of tools and connections. | Tool modifies the wrong resource |
Before a task runs
- The Agent Owner defines purpose, operating permissions, quality expectations, and guardrails.
- The Task Owner defines the objective, acceptance criteria, and authorized grant.
- The Capability Owner validates tools, target resource handling, and access restrictions.
- The Platform Owner supplies isolation, observability, and policy enforcement.
Review the outcome and the system
A Task Owner accepting a useful result does not prove the agent behaves well across all sessions. The Agent Owner should track recurring errors, unsupported claims, failed checks, and regressions over representative evaluations.
Record usage, cost, failures, approvals, tool activity, policy decisions, and final acceptance. Agree retention and visibility before collecting sensitive information. Incident review should identify both the failed control and its responsible owner.
Example: an incident investigation
Alex owns the task and reviews the report. The engineering operations lead owns the investigator’s quality over time. The platform team owns execution and enforcement. The observability team owns the logs connection and search tool. If a tool searches the wrong service, that is a capability issue; if an unauthorized call is permitted, policy enforcement also needs review.