Gate
Gate overview
Gate is the decision layer between an AI agent and the actions it proposes. Before an agent refunds an order, cancels a subscription, sends a message, changes a configuration or deletes a record, it asks Gate. Gate answers in well under a second:
| Answer | Meaning |
|---|---|
allow |
The agent may proceed. |
review |
A person decides. Gate says why, as stable reason codes. |
deny |
Do not do it. |
more_information |
Fetch what missing names and ask again. |
Every answer carries a confidence, the reason codes, risk scores, what is missing, how the proposal was read, and a decision id. The decision is recorded and can be fetched later, with the verdict a person gave on it.
Where it fits
Gate never changes a customer's code and never executes anything. There are two ways to connect:
- Your agent calls Gate. One HTTP call,
POST /v1/gate/evaluate, made by whoever runs the agent, with the proposed action, or simply the text and data the agent has. That is the whole integration, like a payment or a login API. See the quickstart. - Gate comes to where the work is. For a company with no code at all, a Työäly connector reads the source of the work and proposes the actions. Control does this for a Microsoft 365 mailbox.
Two environments
Every organisation has a test and a live environment, each with its own keys, settings and
records. A tyo_test_… key decides on test, a tyo_live_… key on live. Nothing crosses between
them.
Advisory mode first
A new environment starts in advisory mode: Gate decides and records, and the person keeps acting exactly as before. Every morning a digest mail lists the decisions of the day before with three links per decision: I did it, I refused it, I did something else. The agreement report joins those verdicts to Gate's decisions. When the report shows the two agree, Gate's allows can be acted on automatically. How advisory mode works.
What Gate is not
- Not an agent framework and not a model: it decides about actions, it does not perform them or generate them.
- Not a filter on content: it judges a proposed action against the evidence, the limits and the rules of your environment.
- Not a place where your data accumulates: the record keeps the decision, the content only as long as the environment's retention mode says. Trust and data.
Something missing or unclear on this page? Write to hei@tyoaly.fi.