HC05 · Control

Agent Attention Triggers

Purpose System-initiated alerts that interrupt a designated human overseer when agent behavior warrants attention — without the agent having requested it. Generated by a monitoring layer that watches execution patterns, scope boundaries, and decision forks in real time.

Use when An agent is executing autonomously over time, and the system needs to surface conditions to a human overseer who is not actively watching the execution — behavioral anomalies, unauthorized scope requests, unresolvable decision forks, stall conditions, or policy violations.

Do not use For agent-requested approval gates (use HC01), or for surfacing information the user was already expecting to see (use HC03 Agent State Indicators).

Failure mode prevented An agent operating outside expected parameters, touching unintended data, or stalling on an unresolvable decision — with no human informed until something has already gone wrong.

Example Dark Academia

Agent is processing records at 4× its baseline rate

The agent has processed 12,400 records in the past 6 minutes — compared to a 90-day baseline of ~2,800 per 6-minute window. No scope change was requested. The monitoring layer flagged this as a statistical outlier.

Trigger source

Execution rate monitor

Deviation

+340% above baseline

Agent status

Still running

Agent attempted access outside approved data scope

The agent requested read access to billing-archive-2019, which was not included in its approved scope definition. The access request was blocked and the agent has been automatically paused.

Scope boundary

APPROVED billing-current, billing-2023, billing-2022

BLOCKED billing-archive-2019

Agent reached a decision it cannot resolve autonomously

Two valid routing paths exist and the agent's confidence is equal for both (48% / 52%). Proceeding with either option falls below the defined confidence threshold. The agent is waiting.

Path A 48%

Route to standard onboarding queue. Lower friction, slower compliance review.

Path B 52%

Route to expedited review. Higher compliance scrutiny, faster resolution.

Overview

Design rule

Monitoring must be independent of the agent being monitored. Every trigger must name: (1) the specific condition detected, (2) the agent and run context, (3) what the agent is doing right now, and (4) the concrete actions available to the human. No dead-end alerts.

Theming API

Override these CSS custom properties to apply your own visual treatment without touching the component's behavioral contracts.

Shared semantic tokens

Token	Controls
`--rad-surface-card`	Card background
`--rad-color-text-primary`	Title text
`--rad-color-text-secondary`	Description body
`--rad-color-text-meta`	Meta text, context labels
`--rad-radius-card`	Card border radius

Component-scoped tokens

Token	Controls
`--rad-trigger-anomaly-color`	Anomaly left border and header accent
`--rad-trigger-scope-color`	Scope escalation left border and header accent
`--rad-trigger-fork-color`	Decision fork left border and header accent
`--rad-trigger-context-bg`	Context metadata panel background
`--rad-trigger-scope-diff-bg`	Scope diff panel background
`--rad-trigger-approved-color`	"APPROVED" label in scope diff
`--rad-trigger-blocked-color`	"BLOCKED" label in scope diff
`--rad-trigger-value-alert`	Deviation/alert value color
`--rad-trigger-value-ok`	OK status value color

"*Something went sideways and the monitoring system caught it. Surface it clearly. Do not quietly continue.*"

← Previous Consent & Scope Gates Next → Recovery & Override