Build the governance layer for AI agents

At the crux of governing AI is understanding that the risk surface never sits still. Every leap in capability changes what actions are possible, what data is reachable, and how much authority a user can delegate. Policies written for one generation rarely survive contact with the next. Governance has to evolve with the systems it governs.

The shift from chatbots to agents is the most recent illustration. In the chatbot era, the unit of risk was the conversation: govern which apps people used and what they uploaded. Today's autonomous agents take actions, often with live credentials and minimal supervision. The unit of risk is now the action chain: a single run can string together a credential read, a query, and a write to a shared system, with no human in the loop. Governance has to operate on that chain and the real user behind every step.

The detection playbook breaks for the same reason. Signals SOC teams built around humans (off-hours access, bulk reads, sudden volume spikes) fire constantly under agents and almost never matter. The signals that actually matter (an agent accumulating scopes it shouldn't have, a caller drifting outside its trust boundary, a behavioral chain that doesn't match any user's profile) require an engine that learns what normal looks like per entity, not one that alerts on every deviation.

What's needed is a different shape: a unified control plane for agent governance. It should capture every agent action and tie it back to a real user. It should learn behavioral profiles for each agent and user from real session data, propose policy changes for humans to approve, and enforce those policies at runtime by redacting what's unsafe instead of blocking productive work. And it should surface only the findings that actually warrant attention. Policy that evolves with the system it governs, with you in the loop. That's what we're building at Forge, and we need people who want to build it with us.