The problem: agents coded without being thought through
It is easy to open an editor and start chaining LLM calls. It is much harder to stop first and answer the questions that truly determine whether the agent will work and how much it will cost: who is it for? how much autonomy does it have? how will you know if it is right? what does it integrate with? what happens when it fails? Skipping those questions is the fast track to burning tokens on an agent nobody uses.
Just as the Business Model Canvas forces you to think a business through on one page before building it, the Agent Canvas forces you to design the agent before you code it.
The nine blocks
- 01 · Segment / user — who the agent works for and what work it takes off their plate
- 02 · Value proposition — what it does better, cheaper or faster than the alternative
- 03 · Autonomy — where it decides alone and where it stops for human approval
- 04 · Tone / persona — how it communicates and what register it uses with the user
- 05 · KPIs — what is measured to know if it is right; how it is evaluated
- 06 · Integrations — what tools, APIs and data sources it needs
- 07 · Models — which LLM per task: frontier, specialized or local
- 08 · Costs — token budget, cost per task, rented/owned ratio
- 09 · Risks — what can go wrong, and which guardrails and gates contain it
Why it matters before coding
Every block you leave blank is a decision you will make anyway —but by default, mid-development, and usually badly. Filling in the canvas first turns those implicit decisions into explicit ones: the autonomy block tells you where to put human gates, the models and costs blocks stop you from using a frontier model for tasks a local model would do for free, and the risks block makes you design the guardrails before you need them.
It is the natural step before taking the agent to a governed board like Agent Board: first you design the agent in the canvas, then you govern it in production.