Agent governance
Scope agent accountability, permitted actions, evidence, learning authorization and controlled releases.
Agent governance defines an operating contract for each business agent. Record the job, accountable business owner, permitted sources and tools, action boundaries, approval rules and required evidence.
Prepare the operating contract#
| Decision | Review with your team |
|---|---|
| Identity and access | Agent and user identity, tenant boundary, tool scopes, least privilege and revocation |
| Actions | Read, propose and write permissions; spend thresholds; required human approvals |
| Exceptions | Refusal, escalation, interrupted-run recovery and incident responsibilities |
| Evidence | Source references, action and approval records, configuration version and usage |
| Data | Processing region, retention, deletion and access to execution records |
| Cost | Limits, retries, external-tool charges and complete cost per accepted task |
| Release | Acceptance owner, evaluation set, policy regression checks and rollback |
Qualify controls against representative tasks and failure cases in the actual deployment. An illustrative workflow, product page or software test alone does not establish customer control coverage.
Learning authorization#
Execution access does not grant permission to train on data or tool outputs. Specify allowed feedback, data rights, tenant boundaries, retention and compute budget before a learning run. Calibrate agent judgments against expert labels and observable outcomes. Keep governance and access policies outside reward optimization; learning cannot expand agent permissions.
Training can update an eligible model or agent policy. Prompt, retrieval, memory and tool configuration changes have their own evaluation path. Both paths need independent acceptance criteria and versioned release evidence.
Acceptance and ongoing review#
Review the proposed agent against a frozen held-out task set, policy regression cases, latency and full operating cost. Assign responsibility for approval, promotion, rollback and recurring review. Confirm the applicable controls and implementation in the written scope.
Agentic OS provides the operating layer. Forward-deployed delivery covers integration and ongoing ownership. Agent Learning covers reward-based training. Trust review provides the separate security and procurement evidence path.