Skip to content
AM

Agent activity timeline

The agent activity timeline is the operator ledger for AI COO activity. It joins persisted AuditCore events with live pending-action rows and also hosts the current agent policy controls.

What this is

The page lives in Accounting at /agent-activity. It queries agentTimeline with a limit of 100 and supports source and date filters.

Each timeline card shows:

  • Event icon and label.
  • Source badge: Audit for AUDITCORE entries, Action for PENDING_ACTION entries.
  • Tier badge when the payload includes risk_tier or tier.
  • Tool or action label derived from payload.
  • Error, confirmed, or rejected badge when the payload carries that state.
  • Timestamp, relative time, correlation ID prefix, and duration when present.

When you’d use it

  • Review what the agent executed, proposed, approved, rejected, or failed.
  • Filter to pending-action rows or AuditCore rows.
  • Investigate a chain by correlation ID.
  • Tune the tenant agent policy after reviewing behavior.
  • Show a compact operational history to an accountant, auditor, or admin.

Walkthrough

Summary badges. The page counts total, executed, failed, proposed, approved, and rejected entries from the currently loaded result set.

Filters. The source filter supports all, Audit Events, and Pending Actions. The date input sets the since variable for the GraphQL query.

Timeline entries. Event icons are derived from the event type: approved, rejected, proposed/pending, failed, executed, or generic agent activity. Source comes from the joined timeline source, not from free text in the payload.

Correlation IDs. When present, the card shows the first eight characters of correlationId. Multiple events from one agent run can share a correlation ID.

Duration. When the payload includes duration_ms, the card shows the value in milliseconds.

Agent policy card

The policy card is above the timeline. It reads agentPolicy and saves changes through updateAgentPolicy.

Current editable controls:

  • AI Agent Enabled - master toggle for the assistant.
  • Daily T3 Limit - maximum mutating actions per day. Blank means unlimited.
  • Auto-Approve Below (cents) - T3 actions below this amount can skip approval.

The current UI also displays Blocked Tools badges when the policy already contains blocked tools. Blocked tools and tier overrides exist in the GraphQL policy shape, but this operator card does not currently provide editors for those fields.

What “visible to the agent” means

The agent does not receive every catalog tool. Visibility is filtered by the tool catalog, tenant entitlements, tenant features, GL-layer requirements, role and permission checks, persona constraints, and the tenant agent policy.

The policy layers on top of catalog visibility:

  • Disabled agent policy prevents new agent work.
  • Blocked tools are hidden from the agent.
  • Tier overrides can make a tool require stronger gating.
  • Auto-approval and daily T3 limits control whether eligible T3 writes execute immediately or require confirmation.

If the agent does not use a tool you expected, common causes are missing plan access, missing role entitlement, a feature flag, GL-layer restrictions, or policy blocking.

Programmatic access

  • GraphQL (accounting-api): agentTimeline(correlationId, since, until, source, limit) returns timeline rows. agentPolicy reads the current policy. updateAgentPolicy(input) saves policy updates.
  • REST (accounting-api): GET /api/agent-policy returns the policy plus agent access context for the AI COO Worker at chat initialization.
  • AuditCore provenance: use AuditCore provenance APIs for deeper event evidence by event or correlation ID. See The audit log and the change ledger.

The current AI COO MCP catalog does not expose agent-policy or activity-list management tools. Agents use MCP for tool execution; this operator policy and timeline surface is exposed through the Accounting API/UI layer.