Agent Design Canvas
A comprehensive framework for designing AI agents
Think through all the key aspects of agent design including triggers, knowledge, outputs, tools, risks, human collaboration, and success metrics.
About the Agent Design Canvas
The Agent Design Canvas is a practical tool created by the Abundly team to help you systematically design AI agents. Whether you're planning your first agent or refining an existing one, this canvas provides a structured approach to thinking through all the critical aspects.
Perfect for both individual brainstorming and group workshops, the canvas ensures:
- You don't miss any important considerations when designing your AI agent
- You have a structured way of discussing and aligning on the agent within a group
- You can spot weaknesses during the design process instead of wasting time on a setup that would never work
See our agent design canvas below. We have outlined what kind of questions should be answered in each section of the canvas.
Agent Design Canvas
Purpose
• Who would benefit, and how?
• What is the precise use case of the agent? Give examples of how it can be used.
• What is out of scope for the agent in this version?
Impact
• What is the measurable gain to the organization once agent is done?
Triggers
• What event starts the agent?
• Is it manual, scheduled, or event-driven?
Input
• What data or documents are provided?
• What format should inputs be in?
Action Plan (Human & AI)
Create numbered steps showing the workflow, clearly indicating which steps the AI agent performs and where humans are involved. Include review points, approvals, and handoffs between human and AI. Each step should specify who does what.
Example:
1. AI: Analyzes incoming document
2. AI: Extracts key information
3. Human: Reviews and validates findings
4. AI: Generates draft response
5. Human: Approves and sends
Interfaces
• How will human(s) and AI interact (chat, shared doc, approve, joint doc, other)?
• What are the functional requirements on the interface(s)?
Knowledge & State
• Any external systems to connect to?
• What context or domain knowledge is needed?
• Does the agent maintain and modify any states?
Capabilities
• What integrations are needed?
• Other tools needed to perform actions well?
Output
• What are the deliverables?
• What format should they be (templates, examples)?
Success
• What is the definition of done?
• How do we know it became better/worse (evals)?
Example: Manuscript Screening Agent
This is an example of an agent design canvas for an agent which screens incoming manuscripts for a publisher. The agent helps to sort and filter among all the incoming manuscripts in order to find the best ones for a human colleague to review further.
Agent Design Canvas
Purpose
Who benefits: Editorial team saves time by focusing only on high-quality submissions.
Use case: Screen incoming manuscripts to the public submission inbox.
Out of scope: Filtering higher quality submissions.
Impact
Screening efficiency increased by 80%, allowing the editorial team to focus on high-quality manuscripts.
Triggers
Incoming email from the public manuscript submission email address
Input
Each incoming email is accompanied by the email body itself, plus a manuscript and cover letter as attachments
Action Plan (Human & AI)
1. Agent receives email with manuscript and reads the incoming email, cover letter and first 5 pages of the manuscript
2. Agent reviews the content based on quality of writing, and sets a traffic light score (yes, maybe, no) on the email
3. Agent summarizes the recommendation and the motivation behind it, and emails it to the human user
4. Human user reviews the agent's assessment and takes the best manuscripts for further review
Interfaces
Email: Human will receive reports via email from the agent. No further requirements.
Knowledge & State
Agent needs examples of what is great, just-good-enough and not-really-good-enough levels of quality.
Capabilities
Receive email, send email, access/read documents
Output
Agent's review of the incoming manuscripts is aligned with how the human user would evaluate them.
Success
Definition of done: Agent reviews 100% of incoming submissions within 24 hours and provides consistent scoring. Target pass rate: ~40%.
The example above is a fairly straight-forward agent. For more complex workflows, any of these sections could mean a significant amount of time spent for figuring out the right solution.
One common mistake that people make is trying to put everything they can think of into the agent's context. The problem is that this leads to a low signal-to-noise ratio (or a high noise-to-signal ratio), which can lead to a confused agent and often low quality output. Try to the limit the context to what the agent actually needs to know in order to fulfill its duty, and remove anything that is only good-to-know.
Get your own canvas: You can use the interactive HTML version to fill in digitally, or download the PDF version to print and fill in manually during workshops.