DEV Community

Aymen Segni
Aymen Segni

Posted on

Shipmoor: the verification layer for AI agent loops

#automation #ai #programming #productivity

Shipmoor Claim Check EvidencesAgent loops write code far faster than any team can review it, and the validation layer never scaled to match the code generation layer. The bottleneck used to be writing the code. Now it's trusting it to run in production.

Code review is the validation layer the industry relies on today, and it wasn't built for this. A reviewer - human or AI - reads a diff and forms an opinion about whether it looks right. That doesn't verify whether the code actually accomplishes what was prompted. It can't confirm the change aligns with its intended spec, or that the agent's "task completed" claim was checked in a deterministic way.

I'm Aymen - a platform/reliability engineer with 15 years across cloud providers, fintech, DevTools, and e-commerce infra, now working on agentic AI systems. I built Shipmoor to close that gap.

How Claim Check works

Claim Check freezes your acceptance criteria — auto-derived from your prompt or ticket, approved with a single git commit — into atomic, checkable obligations. Each one has to bind to real evidence:

  1. Build passing
  2. Tests running (and catching it if an agent quietly deleted or weakened a test to force a green)
  3. Deterministic scans for known AI code defects
  4. In-depth AI-assisted code review
  5. Before/after state checks

Every piece of evidence is wrapped in an in-toto attestation, and the verdict ships as a SLSA-shaped Verification Summary Attestation (VSA) with a self-digest. Shipmoor checks whether two runs match byte-for-byte, instead of taking a rerun's word for it. A deterministic threshold evaluates the claims on that evidence alone.

AI advises. Evidence decides. Your LLM (bring-your-own-agent) only ever looks at what's left over — the obligations nothing deterministic could bind to a real check. Even then, it gets exactly two moves: block by pointing at a concrete divergence, or say it can't tell. It never gets to grant a pass.

The loop in practice:

  1. Your AI agent makes a code change.
  2. Claim Check tries to bind each obligation to real, internal evidence.
  3. On failure, it returns a concrete reason. Anything not yet verifiable is marked unresolved, not assumed passing. Either way, the agent gets a fix checklist — the exact checks it still needs to satisfy.
  4. Your agent writes the fix, and the system checks it again against the same obligations.
  5. Only when everything is verified does the run come back READY.

Here's a 3-minute terminal walkthrough of the loop: Claim Check Short Demo

It's not just Claim Check

Claim Check is the core of what I've been talking about, but it's not the only thing Shipmoor does:

  • Test Evidence: checks whether your tests actually ran and passed, and catches it if a test was quietly deleted or gutted to force a green build.
  • Agent Harness: gates Claude Code, Cursor, Codex, and others right in the edit loop, so the agent can't call a change "done" while it still has blocking issues.
  • Agent Skills: puts the same checks natively inside Codex, Claude, and Cursor, wherever you're already working.
  • CI Integration: a ready-to-use workflow runs a floor-only gate against local commits (pre-commit) or in CI.
  • Local & private: the review pass runs inside your own coding agent. No model is hosted or called by us. No source code upload.

The scanning layer underneath all of this is always free. No account required to try it:

curl -fsSL https://dl.shipmoor.dev/install.sh | bash
shipmoor scan --changed
Enter fullscreen mode Exit fullscreen mode

The full loop: Claim Check, Test Evidence, Code Review, Agent Harness, and Agent Skills, is the Shipmoor Individual Contributor (IC) plan, $19/dev/month.

Try it

We're opening invite-based beta access: 30 days free, no credit card. You can claim it right away on our website or email me directly at aymen@shipmoor.dev if you want in or want to talk through where it does and doesn't fit your setup.

I'm also recruiting a part-time founding full-stack engineer (equity/shares only). If this problem space is one you think about, reach out to that same address.

More at shipmoor.dev and the engineering blog if you want the deeper design write-ups behind Claim Check and Test Evidence.

Top comments (0)