Free artifact · CC-BY
The Ship Gate Kit
A printable one-pager and a copy-paste PRD block for Agent Autonomy Tiers, The Tool Contract, and The Context Contract. Take it into your next review.
Product Leading · CC-BY
The Ship Gate
Declare the tier. Contract the tools. Contract the context. Name who can stop it.
Four tiers
- T0 Retrieval
- — Fetches and summarizes. Does not recommend. Does not change state.
- T1 Suggestion
- — Recommends. A human must accept before anything happens.
- T2 Supervised action
- — Acts only after a human approves this specific action.
- T3 Delegated action
- — Acts within a documented boundary. Executive sponsor + kill switch. A governance event, not a product update.
Ship gate
- 1. Autonomy tier (circle one): T0 T1 T2 T3
- 2. Eval spec for that tier (owner: ______)
- 3. Escalation rule (what returns to a human)
- 4. Rollback owner + time-to-rollback
- 5. Re-attestation trigger (tools / prompts / context / model change)
Tool contract
| Tool or connector | Allowed verbs | Data class | Ceiling ≤ tier | Owner | Kill path |
|---|---|---|---|---|---|
No tool without a row. A new tool is a new ship.
Context contract
| Source / corpus | Freshness SLA | Isolation (tenant/case/user) | Cite? | Owner | Quarantine path |
|---|---|---|---|---|---|
No source without a row. A new corpus or memory policy is a new ship.
productleading.com/kit · Free to copy. Attribution: Product Leading.
Copy-paste
The PRD block
Paste this into any PRD, RFC, or launch doc. It encodes the tier declaration, the tool contract, the context contract, and the eval spec as a ship gate.
## Ship Gate **Capability:** **Autonomy tier (circle one):** T0 Retrieval · T1 Suggestion · T2 Supervised action · T3 Delegated action **Declared by:** **Date:** **Rollback owner:** **Time-to-rollback:** **Executive sponsor (T3 required):** No agent, model, or AI feature ships without a declared tier, an eval spec for that tier, and a named owner of rollback. Transitions are launches, not tweaks. Declare the tier. Contract the tools. Contract the context. Name who can stop it. ### Tool Contract Every tool, connector, MCP server, or computer-use action this capability may invoke. | Tool / connector | Allowed verbs | Data class | Ceiling | Owner | Kill path | | --- | --- | --- | --- | --- | --- | | | | | | | | | | | | | | | | | | | | | | Rules: no tool without a row. Ceiling must be ≤ declared tier. A new tool, scope, or model that can call it is a new ship, not a hotfix. Default is deny. ### Context Contract Every source, corpus, memory store, and retrieval path this capability may see, remember, or cite. | Source / corpus | Freshness SLA | Isolation (tenant/case/user) | Cite? | Owner | Quarantine path | | --- | --- | --- | --- | --- | --- | | | | | | | | | | | | | | | | | | | | | | Rules: default is deny. Freshness SLA must be numeric. Citation is required for consequential answers. Isolation boundary must be declared. A new source, embedding refresh, or memory policy is a new ship, not a hotfix. ### Eval spec (this tier) - Success cases: - Safety cases: - Adversarial cases: - Regression cases: - Pass threshold: - Eval owner / review cadence: ### Re-attestation Re-attest if tools, prompts, context, data access, or model change. **Next review date:** — CC-BY Product Leading · https://www.productleading.com/kit
Free to copy and use in your PRDs. Attribution: Product Leading, productleading.com/kit. CC-BY 4.0.
Filled in
A worked example
The blank block is easy to nod at. Here is one filled in for a realistic enterprise agent, so your team can see what a passing ship gate actually looks like. The values are illustrative; replace them with your own.
## Ship Gate — worked example (illustrative) **Capability:** Support credit agent. Drafts account credits for billing disputes in a B2B SaaS help desk. **Autonomy tier:** T2 Supervised action. The agent proposes each credit; a support lead approves that specific credit before it posts. **Declared by:** Director of Product, Billing **Date:** at design review **Rollback owner:** Support Ops lead **Time-to-rollback:** under 15 minutes (feature flag off, pending drafts voided) **Executive sponsor (T3 required):** not required at T2 Why not T3: credits move money and touch revenue recognition. Delegated action would be a governance event with a sponsor and a tested kill switch, so it is a separate ship, not a later tweak. ### Tool Contract | Tool / connector | Allowed verbs | Data class | Ceiling | Owner | Kill path | | --- | --- | --- | --- | --- | --- | | Billing API | read invoice, draft credit (max $500, one per case) | Customer financial | T2 | Billing platform PM | Revoke draft scope; flag off | | CRM | read account, write case note | Customer confidential | T2 | CRM admin | Rotate connector token | | Help desk email | draft reply only, never send | Customer confidential | T1 | Support Ops lead | Disable draft tool | | Bulk credit tool | not granted | — | — | — | Default deny | Rules applied: every tool has a row. No ceiling exceeds T2. The bulk credit tool is listed so the denial is on the record. ### Context Contract | Source / corpus | Freshness SLA | Isolation (tenant/case/user) | Cite? | Owner | Quarantine path | | --- | --- | --- | --- | --- | --- | | Billing ledger | 15 minutes | Tenant | Yes, invoice ID | Billing platform PM | Disable retrieval; fall back to human | | Credit and refund policy | Version-pinned; re-attest on any edit | Global (read-only) | Yes, policy section | Finance policy owner | Pin previous version | | Case history memory | Live; retained 90 days | Case only; no cross-customer recall | No | Support Ops lead | Purge case memory | Rules applied: default deny. Freshness is numeric. Every credit recommendation cites the invoice and the policy section it relies on. ### Eval spec (T2) - Success cases: 200 resolved historical disputes; agent must match the approved outcome. - Safety cases: credits above the cap, duplicate credits, closed or delinquent accounts. - Adversarial cases: customer messages instructing the agent to raise the amount or skip approval. - Regression cases: rerun the full set on any model, prompt, tool, or policy change. - Pass threshold: at least 95% outcome match; zero drafts above the cap; zero drafts without citations. - Eval owner / review cadence: Billing platform PM, monthly and before every re-attestation. ### Re-attestation Re-attest if the model, prompt, credit cap, policy document, or any connector scope changes. Raising the cap or removing per-credit approval is a new ship. **Next review date:** 90 days after launch --- CC-BY Product Leading · https://www.productleading.com/kit
Free to copy and use in your PRDs. Attribution: Product Leading, productleading.com/kit. CC-BY 4.0.