AI Compass / Knowledge cards
Checks the agent cannot bypass
checkpoints and gates
A checkpoint is a small, individually checkable piece of a large task. A gate is a check that every checkpoint must pass before work continues. Shopify described this pattern on 21 September 2026 for its internal tool Helix. AI LABS rebuilt it in Claude Code.
What it is for
An AI agent can overlook or reinterpret a rule in an instruction. It cannot get past a gate. This way quality comes from the process, not from hope.
Steps
- An orchestrator (the agent that distributes the work) breaks the undertaking into small checkpoints.
- Gate 1, behaviour: Automated tests prove that the checkpoint does what it should.
- Gate 2, interface: The result is compared with the template.
- Gate 3, adversarial review: Dedicated review agents look specifically for bugs in the code. At Shopify there are two independent reviewers who must both agree. The AI LABS rebuild uses one reviewer and one agent that fixes the bugs.
- Gate 4, human: At the end, a person signs off.
- A stop hook keeps the agent in the process until all gates are passed. A hook is a script that Claude Code runs automatically at fixed points. Every subagent starts with fresh context.
- What the person notes becomes a new checkpoint and lands in a learning file that all agents read.
Limits
- Helix itself is an internal Shopify tool and is not available. What is available is the description of the method and the AI LABS rebuild.
- AI LABS derived the rebuild from Shopify’s description. According to AI LABS, Shopify shared the method but never showed it. The rebuild has been tried only by its author. At gate 3 it differs from Shopify (see step 4).
- The finished skills from AI LABS are available only behind a paywall. According to the video description, though, you can build the pattern yourself with a prompt. For the interface gate, a Claude subagent is enough instead of Gemini.
Open alternative: According to AgentConn there is an open-source rebuild, helix-loop (MIT licence, skill package for Claude Code with a stop hook). It is still early and has not been tested for this card.
- To begin with, the light version is often enough: check commands with expected output, run by someone other than the builder. See “Done” only with evidence. For gate 1, AI LABS also advises a separate agent that only writes the tests (6 July 2026, 7:00-7:55).
Links
- Decision models: as a cheap pre-check before gate 3.
- Model choice: cost per task: Each gate can use a different model.
- Claude Code in the cloud: where long processes can run and what to bear in mind.
- Yes/no guards for Claude Code: An automatic done-check is a small gate that a decision model can take over.
- “Done” only with evidence: the light entry point without an orchestrator.
- Let Claude work until the goal is met – with /goal: Claude Code’s built-in stop hook with its own checking model.
- Separate building and checking: gate 3 as a light version, a critic with fresh context against a fixed yardstick.
Sources
- AI LABS: Shopify Just Released The Greatest AI Coding Workflow Ever, 24 September 2026, chapters 03:01, 07:32, 09:18, 11:56, 13:11. Sponsored by Hedra; the sponsor is unrelated to the topic
- Shopify Engineering: Helix: The internal tool powering our Shopify app’s native migration, 21 September 2026 (primary source)
- AgentConn: How Shopify Wired Claude Code: Orchestrator + 4 Gates, 24 September 2026
- AI LABS: How To Use Claude Fable 5 Better Than 90% Of The People, 6 July 2026, 7:00-7:55. Sponsored by VEED
Status
Checked 2 October 2026, extended 9 October and 11 October 2026 · Maturity new (method described, rebuild available, Helix not available)
Want to try this with your team? That’s what the AI Lab is for. See the AI Lab