During the design process, the AI planning agent helps the thunk's owner or admins create and change a thunk. You describe what you want, the agent drafts a workflow plan, and you refine and test that plan until the thunk is ready for production. At any point you can ask the agent for changes or guidance.
Describe what you want
When you create a new thunk, you describe the workflow you want in a few natural-language sentences. The AI planning agent uses that description to draft a workflow plan: a sequence of steps that carries out the process, plus the workflow state (the properties each work item holds) that the steps read and fill in.
The planning agent automates the first draft, but you can change any part of the plan.
Work with the AI planning agent
The Planning tab shows the plan on the left and the AI planning agent's chat on the right. Tell the agent in the chat what you want changed, and it updates the plan.
To give the plan the full width, select the AI button at the right end of the Planning tab's tab bar (its tooltip reads Hide AI). Show AI brings the chat back. Thunk.AI remembers this choice for each thunk. While the planning agent is generating or updating the plan, its chat stays open so you can follow its progress. A custom-built tool's page has the same button, and it shares the setting with the Planning tab: hiding the AI on one hides it on both.
Refine the design
AI instructions usually start out short and high-level when a thunk is first created. As you test and refine the prototype, they tend to become more detailed and more specific. By the time a thunk goes into production, its AI instructions should be as specific as possible, so that it runs reliably with few errors. See How to Write Effective AI Instructions.
Beyond the steps themselves, you can refine the design in several ways:
Workflow state. Open Work Items in the Run section of the left pane to see all the work items in a table. Move the mouse over a column header to see its column menu (the three vertical dots). Choose Column properties to see and change the definition of that property, Add column to add a new property, or Delete column to remove one. Built-in columns such as Name and Status cannot be deleted. See Workflow Plan for how to define good properties.
Content folders. Describe the content folders you need, and AI agent tasks build them, fill them and index them for search. See Content Folders.
Custom-built tools. Describe the custom AI tools you need, and AI agent tasks build them. See Custom Tools.
Connections and settings. Connect to external business applications, add AI tool libraries, and change the settings that control how the AI agents behave. See Connections to Business Applications and AI Governance.
People. Add the people who work in the thunk and choose what each of them can do. See Users and Access Controls.
Designing a thunk is usually a cycle of authoring, testing and evaluation, and Thunk.AI supports each part of that cycle.
Testing your thunk
Testing a thunk happens in four distinct modes, and a thunk that is ready for production has been through all of them.
Hands-on testing — drive the workflow yourself while you are building it: put the thunk in test mode, re-run a step after changing its instructions, send a test webhook, fire a schedule on demand. See Manual testing.
Diagnostics — after a step has run, open it to see every tool call it made, how long each part took, what it cost, and why it failed. See Analysis of AI behavior.
Regression testing — mark work items as test work items, state what must be true of each one's output, and re-run them whenever anything changes. See Automated tests for workflow steps, and Automated tests for custom tools for the narrower, faster tests that run a single tool.
Design review — a static analysis of the workflow's configuration and instructions that surfaces reliability problems before they cause a failure. See Workflow Review.
Deciding what to test — which inputs, which assertions, and when the suite says you are ready to ship — is covered in Test plan design.