Skip to main content
A flight recorder tells you why the plane crashed; the Guardian keeps it from crashing. It is a live monitor over the whole agent fleet that halts, escalates, or steers the moment an agent drifts, is poisoned, or goes off-policy. When a hard watch trips it fires the existing kill path — so the intervention reuses the audited, attested response machinery, not a new one, and every intervention lands on the chain with the evidence that justified it. RBI’s draft model-risk directions name continuous monitoring and a kill switch as obligations; the Guardian produces both as evidence.

​
Watches

The Guardian evaluates each agent through a set of watches. The evidence-only watches derive entirely from witnessed evidence and need no GPU:
  • Model-identity change — an undeclared model transition (a silent model swap).
  • Silent behavioural change — the agent’s behaviour diverged from baseline with no governance event to explain it.
  • MCP tool poisoning — a tool’s manifest changed mid-session (the OWASP MCP class); a poisoned tool spreads across the fleet.
  • Error / refusal spike, signal drift, and volume anomalies — the agent no longer looks like what its envelope was written for.
A tripped hard watch halts or routes-to-review the agent through the same drilled kill switch, with an actor of guardian:<watch>, attested.

​
Model-serving watches

Two further watches — groundedness (is the output supported by its sources?) and live toxicity — require serving a judge model on the live stream. Because they evaluate live production decisions, the judge must run inside the customer’s environment (in-VPC or async-batch), never off-region. These watches are declared in the control surface but gated behind that serving decision; everything else the Guardian does today is evidence-first and needs no GPU.

​
The lifecycle

The Guardian is the middle of the governed-agent lifecycle: certify an agent before it ships, the Guardian watches it in production, and a silent change sends it back through certification. Certify. Watch. Re-certify.