Gate Routing Map
A gate that only says pass or fail is hiding a decision. Give every case a route.
Built by Sunil Mathew, co-authored with Claude
- What to bring
- One agent you release through automated gates; the list of gate cases or eval suites it runs; for each one, what it checks and how bad it would be if it shipped broken; and the names of any systems that act on the agent’s output, with the teams that own their pipelines.
- What to do
- List your gate cases, give each one a route (Block, Human decides, Check downstream or Track), fill in what each route needs, and score the ten checks.
- What you leave with
- A routing map for every gate case, a list of the cases whose route isn’t ready, and a decision: Ready, Fix first or Hold.
Excel workbook · also opens in Google Sheets and LibreOffice · free
From The Living Craft, a live cohort on designing agentic systems. Explore the cohort
Pass or fail treats every result the same. Agent results are not the same kind of thing.
On an agentic system Sunil worked on at a hyperscaler, automated test cases served as the deployment stage gates. The results couldn’t all get the same treatment. Crucial cases went to a person whose call could stop the release. Others only meant something once a dependent system — a separate service with its own deployment pipeline — had acted on the agent’s output.
A gate that only says pass or fail treats every result the same way. But among an agent’s results:
- some must hold on every run;
- some need a named person’s judgement;
- some only mean something once another system has acted on the agent’s output, and that system may ship from its own pipeline;
- some measure work still in progress and shouldn’t block anything.
Every gate case gets one, and each route needs something before the release.
| Route | Use it when | What the route needs before the release |
|---|---|---|
| Block | A regression case a check can decide on its own | Enough trials, and every trial must pass (pass^k). One green run of a non-deterministic agent overstates its reliability. |
| Human decides | A serious case that no check can decide unambiguously | A named owner, a backup owner, and halt criteria written down before the release. People watching releases get used to green and start reading small anomalies as noise. |
| Check downstream | Another system acts on the agent’s output and ships from its own pipeline | The system and its owning team named; an output contract; the downstream side verifies that contract before either side deploys; a way to switch the new behaviour on together; a named person who can halt. |
| Track | Capability cases, or low-severity cases | An owner who reviews the trend. A drop is worth investigating, not a reason to block. |
The three questions
| # | Question | Tab |
|---|---|---|
| 1 | Which route does each gate case need? | 1 · Gate Inventory |
| 2 | Does each route have what it needs? | 2 · Route Requirements |
| 3 | Is the gate ready to decide releases? | 3 · Checklist |
One question each, and the mistake it catches.
Which route does each gate case need?
The mistake it catches Treating every failed case the same way. Purpose, severity, who acts on the output and whether a check can decide it suggest a route for each case. You can override the suggestion, and the sheet asks you to write down why.
Does each route have what it needs?
The mistake it catches A route on paper with nothing behind it, such as a human gate with no halt criteria, or a Block case that passed one run of a non-deterministic agent. Each row names the first thing missing.
Is the gate ready to decide releases?
The mistake it catches Letting a gate decide releases before its routes are ready. Ten checks, four of them critical, end in one decision and name the next thing to fix.
Four are critical. One No on any of them is a Hold.
Each check is answered Yes, Partial or No, and scores 2, 1 or 0. A No on any check marked critical means Hold. Otherwise 16 or more out of 20, with every critical check answered Yes, is Ready. Anything less is Fix first.
- 01
Critical Every gate case has a route.
- 02
Every route that differs from the suggestion has a written reason.
- 03
Critical Block cases on non-deterministic behaviour run several trials and require every trial to pass.
- 04
Capability cases are tracked, not used to block a release.
- 05
Critical Every Human decides case has a named owner and halt criteria written before the release.
- 06
Every Human decides case has a backup owner, so a release never waits on one person.
- 07
Every Check downstream case names the system that acts on the output and the team that owns its pipeline.
- 08
Critical The agent’s output has a contract that the downstream system verifies before either side deploys.
- 09
New agent behaviour can be switched on together with the downstream system, and switched off without a redeploy.
- 10
Changes to the agent’s output are made in expand-then-contract steps, so the old and new shapes work side by side until the downstream system has moved.
An order-support agent, held.
Illustrative example: an order-support agent whose refund recommendations are carried out by a separate payments service. Not a real system, company or real figures. It is filled in on the workbook’s last tab.
8
3 Block, 1 Human decides, 2 Check downstream and 2 Track. One case was moved off the suggested route, with a written reason.
5
One is “Refund amount within policy cap”. The payments service never verified the contract before deploying.
- Declines refunds on accounts held for fraud review: no halt criteria.
- Resolves complaints that span several orders: no owner.
- Escalation note to the support team: trials that must pass is lower than trials.
- Cites the correct policy clause: fewer trials than the minimum for non-deterministic cases.
Hold
Don’t let this gate decide releases yet. Critical checks 3 and 5 were answered No. Total 12 of 20.
The first thing to fix is check 3. Block cases on non-deterministic behaviour run several trials and require every trial to pass.
The pass rate is still useful. It just can’t tell you who should decide.
You have examined one operating decision. Build the wider reasoning through practical work and review in The Living Craft’s October 2026 cohort: 30 live hours with Sunil Mathew, plus independent work.
Build a working agentic system and develop the reasoning behind it. Explore The Living Craft’s October 2026 cohort: 30 live hours with Sunil Mathew, plus independent work. Apply at https://learning.thelivingcraft.ai/#apply
Published 11 October 2026 · Content CC BY 4.0 · © The Living Craft