An agent should not act until it has passed a check. That is the verdict. Agency Check is the check, and it is free.
Organizations are giving agents authority faster than they are testing it. An agent gets tools, data access, and the ability to change things, send things, and spend money. The deployment meeting covers what the agent can do. It rarely covers whether the agent can be governed. Can it be stopped mid-task? Can anyone see what it did after the fact? Is there a boundary it cannot cross, and who enforces it? These are the questions that decide whether an agent deployment is a controlled system or an open one. Most teams answer them after an incident, when the answers are expensive.
The gap is not a lack of tools. It is a lack of a gate. An agent that has never been tested for readiness is an agent running on trust, and trust is not a control. The organization assumes the harness, the model, and the prompt will behave, because nothing has yet shown otherwise. That assumption holds until the day it does not, and the day it does not is the day the agent does something nobody approved, with tools nobody inventoried, under a policy nobody wrote down.
That is the problem Agency Check was built to solve. Every incident review asks the same questions afterward, and they are always the same four. Could the agent be stopped? Could anyone see what it did? Was there a boundary it could not cross? Was there evidence of what happened, recorded somewhere the agent could not rewrite? If the answers are no, the organization was not running a governed agent. It was running an untested one, and the test it skipped is the test that would have caught it. The cost of skipping it is paid in the incident review, in the remediation, and in the trust lost with the people who depended on the system behaving.
Agency Check closes that gap at the front door. It is a free check at agency-check.app/agency-check. It tests an agent’s readiness and reports what is missing before the agent is allowed to act. The order is the point. Test first, authority second. The check does not ask whether the agent is capable. Capability was never the question. It asks whether the agent is ready to be governed, and it names what is missing when the answer is no.
That ordering is the discipline behind the whole CSI framework, and it has a name: admissibility before execution. Nothing acts until it has been admitted, and admission is a decision made on evidence, not on a demo. A check that reports gaps before authority is granted turns that principle into a step a team can run in an afternoon, instead of a doctrine they agree with in a meeting and forget in production. Governance defines intent. Engineering guarantees reality. The check is the point where the two meet, because it turns the intent to govern into a recorded result before the agent moves.
What to do with it is simple, and the simplicity is deliberate. Run the check on every agent before it acts. Fix what the report says is missing. Run it again. Allow the agent to act only when the report is clean. Treat the result as a gate, not as advice. A gate that can be skipped is not a gate, and a report that gets filed instead of acted on is not evidence. The check is free, so there is no budget argument for skipping it. The only cost is the time to run it, and that time is the cheapest hour in the deployment.
The check is also the start of a sequence, not a one-off event. Agency Check sits in the site navigation under Agentic and MCP, alongside MCP Pro and MCP Services. That placement is the map. The check comes first. It tells you whether the agent is ready. MCP Pro is the governed path for MCP exposure once the agent passes, and MCP Services is the delivery route for teams that want the work done with them. The check is the front door, and the front door is where admission is decided. An agent that fails the check does not get to the next step, and that is the whole point of having steps.
None of this requires a new platform, a new team, or a procurement cycle. The check is free, it runs from the link, and it returns a report of what is missing. The barrier to running it is zero. The barrier to not running it is the incident that follows, and that barrier is paid by the organization, not by the person who skipped the check. Every team that has done an incident review knows the shape of this trade. The hour spent testing before deployment is the hour that would have been spent explaining afterward, and the explanation is always longer.
The launch matters because it changes the default. Before Agency Check, testing an agent’s readiness was a project. It needed someone to design the test, someone to run it, and someone to decide what the result meant. Projects get deferred, and deferred tests become untested deployments. A free check that reports what is missing turns readiness from a project into a step, and a step that costs nothing is a step that gets taken. The organizations that govern their agents well are not the ones with the most controls. They are the ones with a gate that runs before authority is granted, every time, without exception.
Run the check before the next agent acts. If the report is clean, the agent has earned its authority, and the team has evidence to show for it. If the report is not clean, the team knows exactly what to fix, and the agent waits. That is the whole discipline, and it fits in one link.
Sources
- Agency Check, free check. https://agency-check.app/agency-check
- Agency Check info page, Cyber Strategy Institute. https://cyberstrategyinstitute.com/agency-check/
- Release-notes audit, 2026-11-17, gap item 21 (Agency Check launch, no article, no launch date recorded). ~/workspace/seo-review/release-notes-audit-2026-11-17.md