Governance

Controls first, scale second.

Operational agent work needs evals, traces, approvals, least-privilege tools, rollback, and escalation paths before it earns trust.

Define what good looks like before a model run becomes production behavior.

Record provider attempts, failures, lead verdicts, and accepted synthesis without exposing raw transcripts.

Put humans at the points where policy, money, reputation, or irreversible actions enter the workflow.

Keep tool permissions explicit and narrow enough for the workflow's actual needs.

Know which commit, config, or harness change to revert when behavior regresses.

Name the condition that moves work from agent automation back to human review.

What to verify

The workflow has an evaluation or oracle tied to the user-visible promise.
Receipts show which providers ran and what the lead accepted or rejected.
The docs avoid claims that Harness Kit provides RBAC, spend controls, procurement workflows, or incident dashboards.