Candu Actions is currently available to select customers. If you're interested in wizards for your team, contact us.
The agent builder lets you describe the wizard you want, then walks you through configuring it — asking clarifying questions, surfacing what's already in your workspace, and generating a draft you can review before publishing.
It's the fastest path to a working wizard, especially for your first one.
Want the AI to build it for you? Read Build a wizard with AI for a faster path that walks you through configuration via a conversational builder.
This article covers when to use the builder, how the conversation flow works, and what to review before publishing. If you'd rather configure a wizard manually, see Build your first wizard instead.
When to use it
The builder fits four scenarios, each with its own quick-start at the top of the page:
Create a new agent — for a workflow you don't have yet
Clone an agent — to adapt an existing wizard for a related use case without editing the original
Update an agent — to change an existing wizard's behavior, steps, or actions
Improve instructions — when an existing wizard works but isn't reliable enough
You don't have to choose a quick-start. You can just describe what you want and the builder will figure out which path applies.
How the conversation works
Once you start, the builder runs a short, structured conversation before generating anything. Three things happen:
The builder inspects your workspace. It looks at the agents and actions you already have, then tells you. "Here's a quick overview of what you already have: Fix My Agent — diagnoses agent sessions, Action Discovery from Prompt — helps discover API actions..." This prevents you from building a wizard that overlaps with something you already have. If your new idea sounds like an evolution of an existing wizard, the builder will say so.
It asks clarifying questions in structured forms. Rather than free-text back-and-forth, the builder typically uses forms with dropdowns and text fields — "What does your QA sheet look like?" (CSV / Markdown table / Either), "What output do you need?" (free text). You can also answer in plain text in the chat below the form if you prefer.
It summarizes the build plan before generating. Before any wizard config is created, the builder describes what it's about to build — input, scope, output, name, steps, actions needed.
Always review this summary before letting it generate. Push back if anything is wrong:
Before you build it, change the workflow so the agent asks for the session link first, then asks for the expected result.
It's much easier to correct the plan than to undo a generated wizard.
What gets generated
When you confirm, the builder creates the wizard as a draft and shows a summary of what it built — agent title, slug, steps with goals and instructions, actions needed at each step, and completion criteria.
Before you publish
Treat the generated draft as a starting point. Review:
Workflow focus. Is the wizard solving one clear problem, or did it sprawl?
Step structure. Does each step have a clear purpose? Could two be combined? Should one be split?
Action scope per step. Are write actions correctly limited to the steps that need them?
Governance. Are writes set to Confirm required? See Policies for more.
Completion criteria. Does each step have a clear, evaluatable definition of "done"?
If the wizard depends on actions that don't exist yet, the builder may flag this. Set those up first — see Set up actions for your wizard.
Publish the wizard
After you review the draft, the builder will ask whether you want to publish it. Confirm only after you're comfortable making it live.
Once published, the builder confirms the wizard is live and offers a direct link to it. You can also open it from the Agents list to review the full Settings, Placements, and Activity tabs.
The four modes in practice
The same conversational flow handles four different jobs. The framing of your initial prompt determines which path the builder takes.
Create a new agent
Describe the workflow, intended user, and outcome:
Create an agent that helps users troubleshoot why content isn't showing. It should ask for the content ID, check publication status, verify placement rules, and check whether the user is in the expected segment. Multi-step, one diagnostic area per step.
The more specific you are, the better the first draft.
Clone an agent
Reference the source wizard and what should change:
Clone the content troubleshooting agent and create a new version for placement URL troubleshooting. Same diagnostic structure, but focused on URL matching, wildcards, and trailing slashes.
After cloning, the builder will help you rename and adapt — the clone isn't a literal copy.
Update an agent
Reference the target and the change:
Update the content troubleshooting agent. Add a step after the content ID input that checks whether the content is published. If it isn't, explain that and stop the flow.
Review what changed before accepting — particularly whether step order still makes sense and whether action scoping is still tight.
Improve instructions
Reference the step and the behavior you want:
Improve the instructions for step 2 of the content troubleshooting agent. Check whether the user belongs to the target segment. If they aren't, explain clearly and don't continue to placement checks. If the action returns no result, ask the user to confirm the user ID.
Improve is best for steps that "technically work but feel off" — the AI is making the right call most of the time, but missing edge cases or asking the wrong questions.
After you publish
The same loop applies whether you built the wizard manually or with AI:
Test the wizard with realistic scenarios before launch
Check Activity weekly to see how it's performing
Make targeted changes when you spot patterns
See Monitor, test, and improve your wizard for the full workflow.