Agent Design Canvas
Agent Design Canvas, refined by Tree Studios, is a free tool that helps you design an AI agent from two perspectives, value and how, and exports a health-checked build brief you can hand straight to an AI.
Skip the blank page: start from a ready-made agent
Risks, evals and an autonomy ladder are already written. You fill in what is specific to you, like your team, your channels and your numbers, and you have a brief ready to build.
Team chat insights
For product and support teams
A weekly digest of what customers and colleagues say in Slack or Teams.
51 notes already written · 12 blanks to fill in
Use this templateCompetitor monitor
For product and marketing teams
A weekly report of the competitor changes that matter, with sources.
52 notes already written · 13 blanks to fill in
Use this templateRelease notes
For product teams
Outcome-based release notes drafted from finished tickets, approved before publishing.
50 notes already written · 12 blanks to fill in
Use this templateTwo perspectives, one build brief
The canvas starts with the problem, then asks how the agent should work.
Perspective 1: Value
Why the agent should exist, whether it is worth building, and its mission.
Usually filled in by product and business.
Perspective 2: How
How the agent should work, safely, and how you know it does the job.
Usually filled in by the team that builds it.
A file you can hand straight to an AI
The export isn't a copy of the canvas. It's a build brief in markdown, structured the way an AI needs it to start building your agent.
- Goals and context, so the AI makes trade-offs from the problem, not the technology
- Mission, triggers and outputs
- Knowledge, permissions and what the agent must never access
- Hard rules to follow without exception
- Acceptance criteria written from your evals
- Open questions, so the AI asks instead of guessing
invoice-triage-build-brief.md
# Build brief: Invoice Triage > This is a build brief for an AI agent, created with the Agent Design Canvas, refined by Tree Studios. > Always start from goals and context when making trade-offs. > Follow the hard rules without exception. > Items under Open questions are unanswered: ask instead of guessing. > Before building, read the Validation section. Treat warnings and accepted risks as open questions: confirm them with the owner before relying on them. > Never build actions that are not covered by Tools and integrations. ## Key facts - Owner: Head of accounts payable <!-- gate:owner value=Head%20of%20accounts%20payable --> - Changes things outside itself: Yes <!-- gate:acts value=yes --> - Approval before acting: Every time, at the starting level. The Autonomy ladder says when this changes. <!-- gate:approval value=every_time --> - Sensitive data: Financial data, Customer data <!-- gate:data value=financial%2Ccustomer --> - Uses systems or tools: Yes <!-- gate:tools value=yes --> ## Validation Status: Almost ready (0 blockers, 1 warning, 0 accepted risks) Warnings: - W3 Different frequencies are mentioned (hourly, daily and weekly). ## 1. Goals and context …
A health check before you export
Seven quick checks flag what an AI would trip over: no baseline, a hypothesis with blanks left, no rule for what must never happen.
Pick up where you left off
Import a build brief you exported earlier and the canvas comes back exactly as you left it.
Private by design
No account, and nothing is stored on our servers. Your canvas lives in your browser tab until you export it.
Built on Abundly's Agent Design Canvas, with evals and autonomy shaped by Gary Fox's AI Agent Canvas, and a value perspective added in front.
Frequently asked questions
Frame your next agent
in one sitting.
Start from the value, design how it works, and leave with a build brief.
Free. No account needed.