Verified Work Queue Spec (V1)

Use this to design agentic features as operable workflows, not chat demos.

1) Define the Work Object
- Name the “job” (e.g., Refund Review, Access Request, PR Fix, Invoice Match).
- Explicit states: proposed → approved → executing → done / failed / needs-human.
- Attach real artifacts: ticket IDs, PR URLs, record IDs in Salesforce/ServiceNow/Jira/GitHub.

2) Define Allowed Actions (Tooling)
- List each tool/action the system may call (API endpoints or internal commands).
- For each action: read/write/admin classification.
- For each action: required scopes/roles and which identity executes it (user vs service account).

3) Policy Rules (Ship in product)
- Environment boundaries: sandbox vs production.
- Allowlist/denylist actions by workspace/project/customer tier if needed.
- Approval rules by action type (e.g., “write to prod requires approval”).
- Data handling: what gets stored, redacted, retained, exported.

4) Receipt (Immutable event log)
For every step, store:
- Actor + authority: who triggered, who executed, scopes at execution time.
- Intent + constraints: user request, job type, policy version.
- Tool call details: tool name, parameters, timestamps, request IDs.
- Side effects: diffs, created/updated object IDs, links to artifacts.
- Outcome: success/failure, error type, retry count/status, handoff target.

5) Preview & Dry Run
- Provide a plan view or diff before executing write actions.
- Make preview the default path for high-risk steps.
- If no dry run is possible, require approval and show the exact parameters.

6) Reversibility
- Define rollback per action (true rollback or compensating action).
- If rollback is not possible, require explicit approval and a remediation playbook link.
- Provide “undo” affordances where the underlying system supports it.

7) Failure Handling
- Retries per step with idempotency keys.
- Partial completion is normal: record what succeeded and what’s blocked.
- Create a human handoff automatically (ticket/PR/task assignment) with context.

8) Rollout Plan
- Start in read-only mode (suggest/preview).
- Gate write actions behind approval.
- Add limited autopilot only for narrow, reversible actions.
- Instrument with operational metrics you can defend (counts of jobs, failures, handoffs), without relying on model “confidence.”

Ship criteria:
If an on-call engineer can’t answer “what changed, who approved it, and how to undo it” in minutes, don’t ship autopilot. Ship the queue.