Stop an Agent That Is Already Running.
Authorization decides whether an action may happen. Containment is what you have when a chain is already in motion and going wrong. Beacon detects drift against the agent’s own baseline and acts on the whole run, at machine speed.
orchestrator
root of the delegation tree
web_searchget_market_datasend_emailThe Containment Gap
Interception decides one action and is finished. Nothing in that model can reach a run that is already moving.
A Verdict Ends When the Action Does
A per-action decision answers one question and is then finished. It holds no view of the run it belongs to, so when a chain is already in motion and going wrong, there is nothing left to act on.
Detection Hands You a Ticket
Observability tooling raises an alert and waits for a human to triage it. An autonomous agent takes its next action in milliseconds. The gap between the alert and the response is the incident.
The Only Other Option Is Blunt
Without a way to stop one agent, teams shut down whole systems to contain a single misbehaving run, and take the outage rather than the risk.
Four ways to act on a running chain
Allow and deny govern one action. These govern the run, and they reach every agent inside it.
Stop the run
Halt an execution already in progress rather than waiting to refuse its next call. The work stops where it stands, and the reason is recorded with it.
Quarantine the agent
Hold a single agent out of action while it is investigated. It keeps its identity and its history, and it takes no further action until an operator lifts the hold.
Sever the subtree
Collapse everything downstream of the point of failure. A compromised branch cannot keep operating through the sub-agents it spawned, however deep the tree runs.
Revoke authority
Withdraw a grant across the fleet so the same authority cannot be exercised anywhere else, by any agent still holding it.
Effects are enforced at both points Beacon governs: the in-process plugin inside the agent framework and the proxy on the wire. An agent that slips one still meets the other.
Detection That Contains
Knowing an agent has gone wrong is only useful if something happens next.
Drift Measured Against the Agent’s Own Baseline
Behavior is compared to what that agent established for itself and to the plan it declared. Divergence is scored against a threshold rather than judged by a model, so the same behavior yields the same result every time.
Detection That Triggers a Decision
An anomaly score is an input to policy, not a notification. When the score crosses the line policy draws, containment fires at machine speed without waiting for a human to open a queue.
Graduated Response
Containment is not one switch. Policy decides whether a run is stopped, an agent is held, a subtree is collapsed, or a grant is withdrawn everywhere, so the response fits the blast radius.
Release Is Governed Too
Lifting a quarantine requires an operator reason that is written to the audit trail, and the agents held underneath it are released with it. Containment and release are both accountable actions.
What You Get
Grounded in the Principles
Every capability traces back to the 12 non-negotiable principles for Agent Runtime Governance.
Interception Decides. Containment Acts.
Bring your framework. We will show a running agent drift from its declared plan, and you will watch Beacon contain it and seal the record in real time.