Build governed agents directly from your own insights

An agent that is built without reference to a real finding tends to be built for an imagined problem. It watches the wrong thing, asks for access it does not need, or duplicates work someone already does. An agent built from a finding starts with the answers to the hardest design questions: what is it for, what does it need to read, who is responsible for it and who approves what it proposes.

In the Intelligence Platform those answers are designed to come from the graph and from the view you were looking at. The metric is the thing the agent watches. The owner is the person or group the graph resolved. The systems in scope are the ones the finding touched. The evidence behind the finding goes into the agent’s record, so a reviewer can see why it exists.

This is the first half of building from insight: define the agent from what you saw. The second half is governing it, and that begins at the same moment.

Governance is runtime, not paperwork. The agent acts as a known identity, and that identity has permissions scoped to each tool and data source. A policy change alters what the agent can actually do, and does not merely change what a document says it should do.

The agent never has more access than the person who built it, and it never has more than its Trust Profile grants. Retrieved content is treated as untrusted, so a document that tells the agent to ignore its rules cannot override policy. Actions reach the outside world only through typed, approved and audited capabilities. The action gateway that carries them out is on the roadmap, and the design principle is that an agent proposes and the gateway decides.

Autonomy rises by level on the record. An agent that has had its recommendations accepted can move from L1 Assist to L2 Approve, then to L3 Supervise, where it acts within limits and is monitored. Many agents should stay at L2 or L3 permanently. An agent cannot raise its own level.

Review is easier when the agent has a stated reason. A reviewer can ask three plain questions. Does the agent’s scope match what the finding touched? Is each permission in its Trust Profile needed to act on that finding? Is the named owner a person who can be held to account? An agent whose scope is wider than its finding is a design smell, and the review is where it is caught.

The same record supports retirement. When the metric the agent was built to watch has recovered and stayed recovered, the finding that justified the agent no longer holds, and the owner can retire it on the record. Agents that outlive their reason are a large part of how an organisation ends up with an estate it cannot account for.