How Gatehouse works, one capability at a time. Each capability has its own link, so you can send reviewers only the parts they need.
Simulated screens. No customer data appears anywhere.
Features
Gatehouse checks each action before it runs and sends hard calls to a named person. It writes it all to a record you can check and export.
Gatehouse sits between the agent and the systems it uses. It checks each action against your policy before the action runs. Allowed actions go ahead. Anything else is refused, and the refusal is saved to the record.
Each run is written to the trail while it happens: what the agent saw, what it decided, and who made the call. No one writes it by hand. One query returns the full run.
When a step needs judgement, the run pauses and goes to one named approver. The evidence is gathered before they are asked.
An action outside the rules is refused before it reaches your systems. Agents hold only the credentials the job needs, for as long as it needs them. Without a credential, a system is out of reach.
Each approver decision is saved as a labelled example. Before a model, prompt or policy change ships, it is replayed against those examples.
Each run is added to the test suite. Later changes are tested against decisions your team already made.
One file per deployment, written and signed before an agent runs. The agent runs under that signed version only.
The systems and data the agent may reach. Anything outside that scope is refused, and the refusal is recorded.
The role the agent acts in, and what it may spend. The spend limits are written into the manifest, like autopay limit 250.00.
Where the run pauses, and the one approver each case goes to. An override needs two names.
Your team can check each one without taking our word.
In practice
Composite scenarios, labelled as such. Open one to see what happened.
The change goes into the manifest and finance signs it. The next run uses the new limit, because only the signed version is enforced.
The agent tries a refund above its limit. Gatehouse holds it before it reaches your systems and asks the named approver. The hold goes into the record with the rule that caused it.
The agent holds no credential for outbound mail, so nothing is sent. The attempt is refused and saved to the record.
One query returns the full run. The record is sealed, chained by SHA-256 and kept for your retention period.
Simulated, composite scenarios.
