Cross-Agent Audit

June 10, 2026 · View on GitHub

Goal: Answer the accountability test across agent boundaries during incidents and compliance reviews.

When to use

  • Post-incident "which agent did this?"
  • Regulated environment
  • Multi-agent handoffs

Week one (F1)

Read-only audit playbook; no new retention infrastructure required.

Audit query checklist

  1. Time range
  2. Agent ID(s)
  3. Principal (claw vs user)
  4. Tool classes involved
  5. Correlation / trace ID

Minimum evidence chain

User request / schedule
  → agent_id@version
  → principal
  → tool calls (structured)
  → outcome + inbox_id (if gated)

DIY

  • Git log + CI runs for repo agents
  • Export LangSmith traces weekly to team bucket
  • FLEET-STATE.md links to workflow run URLs (this repo's dogfood)

LangSmith Fleet

Native tracing per agent action — combine with agent identity for full clause coverage.

Human gates

  • Audit export containing PII
  • Retention policy changes