Steer active agent populations while they run.

Traditional guardrails and session termination either prevent execution before it starts or destroy context once something goes wrong. PRAE enables security operators to steer active agent populations in real time — fencing cohorts, mutating execution constraints, and changing what agents can do while they are running.

Change the constraint. Keep the execution.

  1. 01 · Active agent

    Mid-session, doing its work: reading, calling tools, reaching data.

  2. 02 · PRAE runtime

    prae-gate at the agent’s tool seam judges each call against the constraints in force.

  3. 03 · Steering decision

    An operator fences a cohort and signs a new constraint with the operator key.

  4. 04 · Updated constraint

    Each targeted agent’s gate verifies the signature and adopts it, with no restart.

  5. 05 · Next execution

    The agent keeps running. Only what the new constraint denies or holds is stopped, and the intervention is on the record.

Govern → Enforce → Prove

In part
  1. 01

    Govern

    Fence a cohort of running agents — by the crown jewel they are touching, environment, team, role, tool or state — and confirm the target, the sessions affected, the new constraint and its expected effect before anything changes.

  2. 02

    Enforce

    The constraint — DATA GUARDIAN · Production Customer Data → DENY, or → REQUIRE APPROVAL — is signed with the operator key. Each targeted agent’s own gate verifies it and adopts it mid-session: later attempts are refused or held, and the agent carries on with the work still allowed.

  3. 03

    Prove

    Each adoption is sealed in a signed ledger row in that agent’s name, so what was adopted, continued, refused or not confirmed is counted from the agents’ own rows, never from what was sent. Removing a steer is sealed as its own event.

Status: in part. On real Claude Code fleets, the fleet control plane changes what every enrolled endpoint in a group may do: a policy version promoted, rolled back or frozen reaches each machine within seconds and applies from its next tool call, with no restart. Fencing a cohort by what it is touching mid-session — a crown jewel, a state, a role — is not yet production-ready: it is demonstrated in PRAE’s local Command Center on a simulated population of Claude Code sessions, where each session’s gate verifies the operator-signed constraint, adopts it mid-session, and seals it in that session’s name. Fleet steering over the Node / TypeScript SDK, the MCP Gateway or cloud runtimes is not built yet.

Fleet control plane →Policy enforcement →The control plane, in context →

  • Not policy enforcement

    Policy enforcement determines what an agent is allowed to do. Live steering lets operators change those constraints while agents are already running.

  • Not session termination

    Ending a session destroys its context. A steered agent keeps running with its context intact; stopping execution is a separate, deliberate action.

  • Not red teaming

    Red teaming tests scenarios before deployment. Steering acts on live agents, during execution.