Shared Inbox HITL

Goal: One place for humans to review, approve, or reject agent actions across the fleet.

When to use

• Multiple agents can take external actions (email, tickets, deploys)
• "Who approved that?" is already a question

Week one (F1)

Approve-only. Agents propose; humans commit. No auto-execute on destructive class.

Model (LangSmith Fleet)

• Claw: editors with edit access review fleet-visible threads
• Assistant: per-user private inbox for sensitive personal tasks

DIY equivalent

• GitHub Issues with label fleet-inbox
• Slack channel #agent-approvals with required reaction before tool continuation
• Linear project "Agent Inbox"

Inbox record shape

## Inbox — 2026-06-10

| ID | Agent | Action | Principal | Status | Trace |
|----|-------|--------|-----------|--------|-------|
| IN-1042 | vendor-intake | Create Linear issue | claw | approved | trace-abc |

Human gates (always)

• Destructive tools
• External customer communication
• Production writes
• Spend above per-action threshold

Failure modes

• FM-03 Inbox Bypass — see failure-modes.md

Verification

Approved actions must link inbox_id in audit log or PR body.