Problem it solves

Once an agent starts running, users lose visibility into what's happening. This creates trust failure when something goes wrong.

When to use

During any autonomous agent run that takes longer than a few seconds or involves more than one discrete action.

When not to use

For single-step, instant actions where the result is immediately visible and there is nothing meaningful to monitor.

Governing principle

The monitoring layer is independent of the agent's self-reporting. If an agent can misreport its own state, the monitoring layer must catch it.

Required Components

Interaction Flow

1

Execution begins

Agent state transitions to RUNNING. The Agent State Indicator surface becomes active.

2

Continuous state updates

The surface shows what the agent is doing right now, what step it's on, and estimated time to completion.

3

Monitoring layer runs in parallel

Independent of agent self-reporting, the monitoring layer watches for behavioral drift, scope escalation, and unresolvable decision forks.

4

Anomaly detected

If the monitoring layer detects a problem, an Agent Attention Trigger surfaces the issue immediately — even if the agent hasn't reported a problem.

5

Human intervention point

The user can pause, redirect, or abort the run at any point. The current state is saved before any intervention takes effect.

Governance requirements

Agent state surfaces must be real-time or near-real-time. Polling intervals over 30 seconds are not acceptable for high-stakes runs. State changes must be logged.

Accessibility notes

State changes must be announced via ARIA live regions. Use role="status" for routine updates and role="alert" for attention triggers. Do not rely on color alone to communicate state.