AI / Automation / Prototype
Agent Canvas
A lightweight way to describe an AI agent before discussing frameworks or models: what it is for, what it knows, what it may do, when it must stop and how success is measured.
Interactive prototype
Define the operating contract.
Mark each area as defined. The goal is not a perfect specification — it is to expose missing decisions early.
Agent design starts before the model.
It is easy to begin with implementation questions: which model, which orchestration framework, which tools, which vector database? Those decisions matter — but they are downstream.
The Agent Canvas moves the first discussion toward the operating model. If the purpose, permissions and stop conditions are ambiguous, a more capable model will usually amplify the ambiguity rather than solve it.
A useful design test
Could a new team member understand what the agent is allowed to decide without reading the code?
If not, the operating contract probably still exists only implicitly inside prompts, integrations and developer assumptions.
Autonomy is a spectrum.
The same use case can often be designed at different levels: observe, recommend, prepare, execute with approval, or execute autonomously. The canvas makes that choice visible so teams can start with the least autonomy required to create value.
What I would add next
- Risk level: attach higher control requirements to more consequential actions.
- Data scope: explicitly list sensitive sources and retention expectations.
- Test scenarios: define normal, ambiguous and adversarial situations before rollout.
- Observability: specify which decisions and actions must remain explainable after execution.