The farther work proceeds between human interactions, the more it depends on feedback the system can act on. Tests, reviews, and access controls help determine whether a change can continue, needs correction, or raises a decision the team hasn't yet made.
Those controls are most effective when they help work proceed correctly, rather than collecting problems for someone to discover at the end.
The check needs to be able to disagree
An agent's explanation of success is one account of what happened. Tests, scanner results, review findings, and the actual changes provide other ways to examine it. They answer different questions, and none covers every aspect of correctness.
A test written alongside an implementation can repeat the same misunderstanding. Independent expectations matter: the requested behavior, an existing regression case, or a reviewer who can question the approach. More reviewer agents don't automatically provide independence if they all accept the same incorrect assumption.
The relevant checks depend on the work. A formatting change and an authorization change warrant different attention. What should remain consistent is the connection between the intended result and the checks used to accept it. A passing pipeline can't settle a question the pipeline never examined.
Instructions and enforcement do different jobs
Instructions explain how an agent should work. Permissions, protected resources, and enforced checks determine what actions the surrounding system allows. Asking an agent to respect a restriction doesn't provide the same control as preventing the action.
The people or agents proposing a change shouldn't also have unrestricted authority to weaken the rules used to accept it. A failing test may legitimately need revision, but removing it to obtain a passing result deserves a different decision from fixing the implementation. Protected changes with independent review explains how that independence applies to policy and automation as well as application code.
Human oversight remains important without requiring a person to approve every routine action. The team can decide which conditions permit work to continue and which situations warrant attention. A request for review should explain the decision needed, the checks completed, and any uncertainty that could affect the answer.
The workflow needs to withstand a mistaken action
A review before merge can't protect a credential that an agent exposed while investigating the change. The working environment matters before the first edit: which files it can reach, which services it can contact, and what its tools allow. Limiting that access reduces the damage a misunderstood request or malicious instruction can cause, even when the agent doesn't recognize the problem.
Control also needs to continue after acceptance. The release process should deliver the artifact that passed the checks, and the team should be able to connect it to the work that produced it. If something goes wrong, stopping an agent is only part of the response; someone still needs to find its changes, account for work already running elsewhere, and restore the affected service.
Evaluation continues after the demonstration
Evaluating an agent and checking its output are related but different activities. Representative cases help a team learn how a capability behaves before relying on it more widely. Checks on actual work help decide whether a particular result is acceptable. Both can miss conditions that appear later in production.
Review findings, unsuccessful attempts, and incidents can improve the next evaluation set or the checks applied during delivery. A change to a model, tool, or instruction may alter behavior even if the task name stays the same. Revisiting representative work gives the team a way to assess that change without assuming either that every update is harmless or that all prior learning has become irrelevant.
Further Reading
- Building feedback and verification into delivery explains how to choose evaluation cases, check actual work, and learn from recurring failures.
- Choosing and revisiting workflow autonomy helps decide what can proceed without another human decision and what should bring that work back for review.
- Security threats in agentic systems describes threats that access controls, review, and tool design need to account for.
- Coordinating human and agent work across delivery explains how access, approvals, and stopping work fit into the way people and agents collaborate.