One page tool · AI Agent Canvas
Design one agent before you build a system
A working canvas for deciding what an agent does, what it may touch, and how you prove it works.
Executives deciding what to automateSmall businesses building a first agentProduct teams before they open a platform
An AI agent system is a group of agents that share identity, data and an audit trail. This canvas gives you the single agent view. Get one agent right, then the system has something to scale.
Why use it
Three things you leave with
01
A job with a clear result and a clear stop.
If the agent’s title sounds like a real job, the scope is clear enough to build.
02
A boundary that limits the worst mistake.
Permissions, approval gates, a fallback and a kill switch, set before the first run.
03
A record that earns the agent more authority.
Every run leaves a receipt. Trust measured as a rate, not assumed at launch.
3 in 4
Roughly three quarters of multi agent failures trace back to design and specification, not to the model. A stronger model does not fix a weak system around it. The discipline lives here, before the build.
The canvas
Six areas, numbered in the order you fill them
AI Agent Canvas
BUSINESS DESIGN LABS
1 JOB
What result does this agent own?
BUSINESS RESULT
The result, and the output that proves it.
TRIGGER AND DONE
SUCCESS METRIC
NEVER LIST
2 LIMITS
Who is acting, and what may it do?
OWN IDENTITY
PERMISSION LEVEL
APPROVAL GATES
A check outside the agent authorises the action. Anything irreversible.
FALLBACK AND KILL SWITCH
3 EQUIP
What does it need to do the work?
MODEL AND PROMPT
TOOLS
MEMORY
Decisions, outputs, failures and open items, stored outside the conversation.
DATA AND RULES
Every dependency named with its source. A short map, not a manual.
4 RUN · the work loop
SENSE
REASON
ACT
OBSERVE
↻ Repeats until a stop rule fires: done, unsafe, or unclear. Retry cap. Budget. Then hand to a human.
5 PROVE
How do you know it worked?
VERIFY AGAINST CRITERIA
LOG AND AUDIT
Action, evidence, cost and rollback point, per run.
TRUST RATE
Share approved without edits. Each failure updates a guide, a test, a tool or a rule.
6 SCALE
One agent to a system. Cross only after areas 1 to 5 work.
GROUP AGENTS AROUND ONE RESULT
ROUTE WORK BY RISK
SHARED SUPPORT: IDENTITY, DATA, AUDIT
CHECK THE WHOLE: RESULT, COST, DEMAND
CoreBoundaryCapabilitiesSystemThe colours show which layer of the agent each area belongs to.
CORE · AREA 1
Identity and job
Who is acting and the result it owns. Everything attaches here.
BOUNDARY · AREAS 2 AND 5
Bounded autonomy
What it may touch, where a human signs off, and how you stop it.
CAPABILITIES · AREAS 3 AND 4
The parts that power the loop
Model, prompt, tools, memory and orchestration. Sense, reason, act, observe.
SYSTEM · AREA 6
The world it runs in
Shared identity, approved data, audit trail and live checks across every agent.
How to use it
Six moves. One decision each.
Finish the thinking before you pick a tool.
MOVE 1 · AREA 1
Define
Name the result, the trigger, the finish, and what it must never do.
Decide: does the title sound like a real job?
MOVE 2 · AREA 2
Permit
Give it an identity of its own. Set the permission level and the gates. A check outside the agent authorises risky actions.
Decide: what is the worst mistake, and does it reverse in ten minutes?
MOVE 3 · AREA 3
Equip
Choose the model and prompt. Start with three to five tools, read only first. Keep decisions and state outside the conversation.
Decide: is every dependency named with its source?
MOVE 4 · AREA 4
Run
Read the case, make sense of it, choose the next allowed step, act, observe. Repeat.
Decide: what fires the stop, and how many tries does it get?
MOVE 5 · AREA 5
Prove
Verify the output against the criteria. Keep a receipt per run. Track the rate approved without edits.
Decide: has it earned the next rung?
MOVE 6 · AREA 6
Scale
Only when one job becomes many. Group agents around one result. Route by risk. Share identity, data and audit.
Decide: does the job split, or does one agent still do it?
THE SCOPING QUESTION
If this agent takes the wrong action here, what happens and who fixes it?
Ask it at every move. It prevents most incidents.
Earn autonomy
Autonomy is a dial, not a switch
Every agent starts at the bottom. It climbs by track record, not by a good demo. The right column is what must be true before the next rung.
4
GOVERNEDExtended authorityActs across a wider scope, still inside permissions, audit and the kill switch.
GATEReversibility and blast radius stay acceptable.
3
MEDIUM RISKSupervised actionTakes actions within defined rules. Humans review the higher risk cases only.
GATESustained reliability on the routine path.
2
LOW RISKDraft onlyCreates emails, summaries and records. A human approves before anything is sent or saved.
GATEThirty days or more at 85 per cent approval, with minimal edits.
1
MINIMAL RISKRead onlyViews records, retrieves data, reads documents. Changes nothing in any system.
GATETwo to four weeks of accurate output with few false positives.
The principles
Seven rules that hold at any size
P1
Define the job first
One result, one trigger, one output, one handoff. Precision in the brief is precision in behaviour.
SCOPE
P2
Bound autonomy, then earn it
Freedom to act inside clear limits. Start read only. Widen authority on track record. Anything irreversible keeps a human in the loop.
CONTROL
P3
Prefer the simplest architecture
One disciplined agent beats a fragile crowd. Add a second only when one job splits into two.
STRUCTURE
P4
Store only what earns its place
For every memory, name the next task it improves. If you cannot, do not store it.
MEMORY
P5
Close every loop to a real system
Output in a chat window creates work. Output that updates a record removes it.
INTERFACE
P6
Prove it, then improve it
Set a baseline. Log every action. Repair the class of failure, not the one output: each incident updates a guide, a test, a tool or a rule.
ASSURANCE
P7
Keep the brain, the hands and the record apart
The model reasons. Tools act inside a boundary. An append only record holds what happened. If one fails, the other two survive.
STRUCTURE
When it breaks
Find the broken part before buying a bigger model
Every agent is built from the same seven parts. When one underperforms, the symptom points to the part.
PART
WHAT IT DOES
BREAKS AS
Model
Reads input, reasons, generates output. Use the right model per step, not the best model everywhere.
Fluent but wrong, vague, or off task.
System prompt
Role, context, rules, format, examples, stop condition. Two to five examples lift accuracy sharply.
Same input, different output. Scope drift.
Tools
Read from and act on real systems. Start with three to five, read only first. Add write tools one at a time.
Confident claims it cannot verify.
Memory
Carries state across steps and sessions, outside the conversation. Decisions, outputs, failures, open items.
Repeats questions. No continuity.
Decision loop
Drives multi step work forward and knows when to stop. A retry cap and a budget, then a human.
Answers once and stalls.
Orchestration
Coordinates the parts into a workflow. What runs first, next, last, and what makes it pause or retry.
Works in test, misbehaves in production.
Interfaces
Connect inputs and outputs to real work. The exit must change something in a system people use.
Great output that lands nowhere.
Shared support
Wraps all seven. A verified identity per agent, then approved data, safe connections, audit, live checks and a versioned model service.
No system works at scale.
FREE DOWNLOAD
Download the AI Agent Canvas
Plus a fortnightly email worth opening. Practical insights, no filler.