What is Action-Time Monitor?

Why It Matters

A pre-action gate can only evaluate what is visible before execution: the proposed call, its arguments, its target. Some violations only appear while the tool runs. An agent granted a read starts a write. A call declared read-only triggers a mutation through a side effect. An action declared for seconds runs for minutes. The action-time monitor exists for the class of violations that pass the gate honestly and go wrong during execution, and unlike every other enforcement point it can interrupt the action mid-run.

What It Watches

Scope creep. The executed action exceeds the approved scope: a directory read that becomes a directory write, a query that mutates, a fetch that fans out into calls nobody proposed.

Side effects. The action produces effects outside its declared contract: network egress from a call described as local, writes to resources adjacent to the approved target.

Timeout conditions. The action runs past its declared duration, which is how a stalled call and a quiet runaway both surface before they become a spend or stability incident.

Where It Breaks

A monitor without an interrupt path is a logger with a better name. If detecting a scope violation mid-execution cannot stop the action, the monitor produces evidence for the postmortem rather than protection during the run. The second failure is checking each call in isolation: an agent that stays inside scope on every call while the run as a whole drifts needs pattern tracking across calls, which is what the wider trajectory is for.

How Flytebit Handles It

In the systems we ship, action-time monitoring is wired to an interrupt path: a scope violation mid-execution halts the action and routes through the same safe-stop machinery as the kill switch. The monitor is one of four enforcement points in Governing Agentic AI, and the engagement version is our AI governance and risk work.