The Agent Canvas
Eight questions to answer before you build anything.
One sheet, eight questions, before you write any code
The Agent Canvas was adapted by Dr Mark Bloomfield from Microsoft's agent design canvas, MongoDB's agentic AI canvas, and Abundly's work — reduced to the eight questions that most matter for a product and engineering team shipping into production.
If you cannot fill cells 5–7 with confidence, go back to cell 1.
1. Purpose — what job is this agent hired to do?
Written as a Job-To-Be-Done: "As a [user], I need to [job], so that [outcome]." Include what the agent is explicitly NOT for. Scope creep kills agent projects.
2. Triggers — what starts it working?
User request, scheduled event, system event, delegation from another agent, time of day, phase of the moon? Be specific about the primary trigger. Secondary triggers imply multiple product flows and a sign you might want a more defined Purpose.
3. Tools & Capabilities — what can it connect to and use?
List each tool and risk-rate it:
- Low risk — read-only (queries, searches, lookups)
- Medium risk — write/modify (create a ticket, update a record)
- High risk — irreversible or high-blast-radius (send an email, make a payment, delete data)
More tools does not mean a better outcome. The agent's action space is defined, and to some extent controlled, by this list.
4. Knowledge & Context — what does it know and remember?
System prompt, retrieved documents, conversation history, persistent memory.
Who controls what it sees? Where is the data stored?
5. Authority Level — what level of autonomy does it have?
Five tiers, earned, not granted. See the authority levels lesson.
6. Guardrails — what is it explicitly prohibited from doing?
Hard boundaries (code-enforced) vs soft (prompt-based). Prompt-only guardrails are not enough — 34% of safety incidents come from soft-only enforcement.
A mixture of fast deterministic and model based guidelines.
7. Human Checkpoints — where must a person approve or intervene?
- Human-in-the-Loop (approves each action)
- Human-on-the-Loop (monitors real time)
- Human-out-of-the-Loop (audits after the action)
Match oversight to decision risk.
8. Success Metrics — how do you know it's working?
Task completion rate. Tool-call accuracy. Groundedness. Cost per run. Latency. User satisfaction. You need to define these metrics, and what they mean for agent operation, before launch, not after.
Interactive canvas
You can fill in a canvas right now. It autosaves to your browser and you can export or share it.
Where this leads
The Agent Canvas is the design artefact; the System Prompt is the build artefact; the REMIT worksheet is the governance artefact. Cells 1–4 of the canvas map to the System Prompt. Cells 5–8 map to REMIT (Envelope, Monitoring, Trust). Done well, the three artefacts are three views of the same agent.