Agent Design Canvas
About this resource
Most agent projects fail at scoping rather than at implementation. The goal was stated loosely, the boundaries were never written down, and nobody specified what should happen when the agent encounters something it was not designed for. All three are cheap to fix on paper and expensive to fix in production.
This landscape canvas walks a team through scoping an AI agent before anything is built: the single goal, the explicit non-goals, the tools and data it needs, where a human must be in the loop, what should happen at each anticipated failure, and how you will know whether it worked. The non-goals section is the one teams skip and the one that most reliably prevents scope drift.
The field teams skip. Writing down what the agent should never do prevents more problems than any amount of specifying what it should.
How to use it
Complete it before anything is built, with the person who owns the process in the room rather than only the person building it. Spend the most time on non-goals and failure modes: those two sections are where the disagreements surface, and surfacing them on a worksheet costs an hour while surfacing them in production costs considerably more.
Inside the download
Deliberately singular, because multi-purpose agents drift.
What it must never do, which is the field that prevents the most trouble.
Exactly what access it needs, which becomes your permission scope.
Where a person must be involved and what they are actually deciding.
Common questions
Start with one goal and write the non-goals explicitly. Then list the tools and data it needs, mark where a human must be in the loop, decide what should happen at each anticipated failure, and define how you will judge whether it worked. Scoping is where agent projects succeed or fail.
Anything irreversible without human approval, anything financially material beyond a set limit, anything that communicates externally under your organization’s name without review, and anything outside the single goal it was designed for. Writing these down is the point of the exercise.
Something written, yes, though a one-page canvas is usually sufficient and far more likely to be completed than a formal specification. The value is not the artifact. It is that scoping forces disagreements to surface while they are still cheap.
Free. No email list, no follow-up sequence.
More from the resource library, chosen because they pair well with this one.
Six controls for software that acts rather than advises, including the one most people never test.
What actually makes something an agent, why it changes the accountability question, and what to ask a vendor selling one.
What has to be true before an AI tool touches a live workflow, including how to roll it back.
Tell us a little about yourself and this document is yours. We don't have an email list. We collect this so we know who's finding our resources useful, and that's it.
You're downloading
this document
Your information is kept confidential and never shared with anyone.
