Tools

Try the method before you talk to anybody

The same questions a design review would ask, in a form you can work through on your own. No sign-in, no email, no score — and nothing you type about your architecture leaves your browser.

Open the agent design check Read the guides

Ungated. There is no lead record created by using anything on this page.

Available now

The agent design check

Browser checklist

Agent design check

19 questions across 7 areas of an agentic design. Answer yes, no, or not yet defined — and get back what is still open, why it matters, and what to do about each one in order.

Open it

What it asks about

Purpose and boundary What the system is for, where it stops, and whether it needed to be an agent at all.
Evidence What must be established before it acts, and whether that can be checked afterwards.
Tool permissions What each tool may do, and where that limit is actually enforced.
External actions What happens when the system does something the world can see and you cannot take back.
Evaluation What you can see after a run, and what the system is checked against.
Uncertainty and recovery What it does when it cannot establish what it needs, and how a person takes over.
Ownership Who is answerable for the behaviour, and what happens when it changes.

What it costs Nothing, and it does not ask who you are

Where your answers go Nowhere. They are held in the page and disappear when you close the tab

What you get back Next steps in priority order, each one written beside its question in the rubric — plus a text file or a printout to take away

What it will not do Score you, grade you, or tell you a system is ready for production

How to read the output

A checklist, not an assessment

There is no number at the end. A readiness percentage would invite you to improve the percentage, and the questions worth answering here are not the ones that move a figure — they are the ones nobody on the team can currently answer.

Every sentence the checker can say is written down in advance, next to the question it belongs to. No model reads your answers, ranks them or writes advice about a system it has never seen. You can read the whole rubric before you answer anything, which is the only reason its output is worth trusting.

Marking every box addressed does not produce a pass. It produces a request for the artefact that would show each claim — the tool list with its scopes, one run's evidence references, the case that came from a real failure. That is what a review asks for, so it is what this asks for.

Not published

Architecture review skill

The roadmap plans a second tool: the same method packaged as an architecture-review skill for Claude and for Codex, reviewing only material you select and returning referenced observations, questions, limitations and next actions — never executing or modifying your code.

It is not built, and this is not a link. It is sequenced behind approval of the rubric the checker above uses, because the skill would package that same rubric and there is no point versioning a method twice. Each host also has to be verified on its own; no compatibility claim will be made across the two before it has been tested on both.

If a question here lands

Pillar guide

Written 11 September 2026.