Agents and the Inbox

How agents work, and why nothing changes without your approval.

datajoi has several agents, each with a narrow job: ingestion, quality (test authoring), transform, knowledge, semantic, and the analysis agent that answers questions. They share one rule: an agent proposes, a person decides.

#The proposal

Every consequential change is a proposal with a type (for example a schema contract, a test suite, a model, a knowledge document or a semantic term), a diff, a rationale, the agent's confidence and the route it will take. You can approve, edit then approve, reject or defer, and add a rationale.

#Routing and earned trust

Some action types are always human: a schema contract that decides what enters your lake never auto-applies. For others, an agent can earn an auto lane after a streak of approvals (five by default). You can see each action type's lane and streak in Settings, and revoke trust at any time. Until a lane is earned, every change waits for you.

#What an agent can and cannot do

• It can read your workspace metadata and, through the guarded read path, your data.
• It cannot run arbitrary SQL writes, call tools you have not listed in policy, or touch another workspace.
• Its output is validated (types, SQL guard, name rules) before it is even shown to you.

See trust and policy for the policy file and responsibilities.