Map the system
Start with the Agent Graph Canvas. Name the outcome, identify loops and groups, and connect them with directed, labeled handoffs. Trace branches, merges, retries, and human boundaries.
Agent design · Paper & digital
AgentGraphs
Design an agent as a graph of focused loops.
AgentGraphs is a design toolkit for making agentic systems understandable and discussable. Each loop has a defined input, one bounded job, and a clear output. Connected loops show how the whole system works—and where people, decisions, and controls belong.
Visit AgentGraphsThe core workflow
From the whole system
to one bounded job.
Design each loop
Use the Regular Loop Canvas to define Input → one bounded job → Output. Add the models, tools, memory, orchestration, and controls that support the job.
Open up the detail
Use the Detailed Loop Canvas to explore Plan → Execute → Evaluate. Specify what happens at each step, inspect results, and decide when to continue, revise, or finish.
Two ways to work
On the table.
On a shared canvas.
Printable workshop kit
The paper toolkit includes loop and graph canvases, card decks, a quick-start guide, a glossary, and worked examples. Cards make the resources, actions, and constraints of each loop tangible.
Collaborative application
The digital app lets participants join a shared project, arrange loops and connections, and work with cards. It can generate a design brief and reflection questions, with Markdown downloads.
When to use it
Design, inspect, and discuss.
Use AgentGraphs before implementation to work through an agent’s structure, or after implementation to reconstruct how an existing system behaves. Its graph and loop views support conversations about handoffs, responsibilities, and the boundaries of autonomous action.
AgentGraphs helps describe and reason about a system. The toolkit and design app do not themselves execute the agents you map.
Ask about using AgentGraphs in a workshop ↗