Skip to content

Steps (Agent Process) ​

Asha runs customer operations at Acme Diagnostics, a lab network that offers home sample collection. When Priya messages to book a blood test, the conversation can't be left entirely to free-form chat. It has to follow a set order: help her choose a test, check that a phlebotomist serves her area, collect her address, and only then book a slot.

Steps let you attach that ordered process directly to an AI Agent. The agent works through the steps one by one, using its own personality, model, knowledge and integrations throughout.

An AI Agent with a Steps chain: Ask agent, Tool, Render View, Collect Info.

Why use Steps? ​

On its own, an AI Agent decides for itself what to do next, guided by its personality and tools. That's ideal for open questions. For a transaction with a required order (a booking, a checkout, an application), you often want guarantees instead:

  1. One brain. Every step uses the parent agent's Personality, AI Model, Knowledge and Integrations. You configure them once, on the agent.
  2. Tools that always run. A Tool step calls one integration tool with the arguments you map. The model can't skip it or make up its result.
  3. A fixed order. The agent moves to the next step only when the current one is done.

The Steps port ​

Every AI Agent has a Steps port along its bottom edge, next to Personality, AI Model, Integration, Knowledge, Action and Operator. Click its + to add the first step. Then click the + on the right of each step to add the next one.

Once an agent has steps, its card reads Agent + steps — runs its workflow, and the steps are what the agent does: each incoming message is handled by the step the process has reached.

The Add a step picker, listing the node types allowed inside Steps.

What you can add as a step ​

StepFound underPurpose
Ask agentAgent processThe agent does one task, such as answering, recommending or asking a question.
ToolAgent processCalls exactly one integration tool with mapped arguments.
Collect InfoDataAsks the customer for specific fields. See Collect Info.
Render ViewOutputShows the customer choices, cards or a checklist in Web Chat. See Render View.
Condition (IF/ELSE)CoreBranches the process. See Condition (IF/ELSE).
Human HandoffCorePauses the process for your team. See Human Handoff.

Ask agent ​

An Ask agent step gives the agent one task. It has no ports of its own: it uses the parent agent's brain.

The Ask agent panel.

Step ​

FieldWhat it doesOptionsDefaultShown when
What should the agent do here?The task for this step — the agent uses its persona/model/tools (from the node) to do just this. e.g. "Write a 3-line visit brief: best arrival time, one must-do, one thing to avoid."
Save reply asVariable name for the agent's reply (downstream steps read {{ name }}). Defaults to ai_response.
Wait for the user's replyAsk the user and PAUSE here until their next message, then continue (for collecting a mobile / OTP / address across turns). Off = do the task and continue in the same turn.false
Advance when this tool succeedsDETERMINISTIC hand-off (use with "Wait for the user's reply" ON): the name of a tool whose SUCCESS marks this step done — the flow moves to the next step the moment that tool succeeds, without waiting for the model to decide. Use for reliable multi-turn steps with small models (e.g. advance on "verify_otp", "set_address", "add_test", "book"). Leave blank ONLY if no tool marks this step done: the step then advances solely when the model decides it is finished, which small models often never do — the step keeps chatting instead of handing off (publish warns about this).

Always fill in "Advance when this tool succeeds" for multi-message steps

Without it, a waiting step moves on only when the model decides the task is finished. Smaller models often never decide that, so the step keeps chatting. Activation warns you about it.

Tool ​

A Tool step calls one tool from the agent's connected integrations, every time the process reaches it.

The Tool panel: tool check_serviceability, argument pincode from a variable, result saved as serviceability.

Tool ​

FieldWhat it doesOptionsDefaultShown when
ToolThe MCP tool to call — one exposed by the agent's connected MCP servers (e.g. create_booking, add_to_cart).
ArgumentsFill each tool argument from a {{ variable }} or a fixed value.
Save result asVariable name for the tool result (downstream steps read {{ name }}).
Result fieldOptional dot-path into the result to store instead of the whole object (e.g. booking_id).
Fail early on errorStop the whole process if this tool errors (e.g. validate a park before booking). Off = pass the error to the next step.false

Click + Add argument for each Arguments input, then choose From a variable (for example {{ pincode }}) or Fixed value.

To branch on how a Tool step went, add a Condition with exit_reason == "tool_step_error" (or tool_step_completed). A Condition can't branch on a Tool step's saved result directly — activation refuses it — so branch on exit_reason.

Withheld tools are refused, not called

If the tool's description was flagged as a possible prompt injection and an admin hasn't approved it yet, the Tool step doesn't call it. It fails with "Tool '…' is withheld: its description was flagged as a possible prompt injection. An admin can approve it in Integrations (MCP Registry)." If another connected integration offers a tool with the same name that isn't withheld, that one is called instead. See MCP Server.

Worked example: Acme Diagnostics home collection ​

  1. Ask agent: choose a test. Task: Help the customer choose a test package based on their symptoms or prescription. Wait for the user's reply ticked, Advance when this tool succeeds set to select_package.
  2. Collect Info: address. Fields pincode (String, required, pattern ^[1-9][0-9]{5}$) and city (String, required), with Confirmation Required ticked.
  3. Tool: check serviceability. Tool check_serviceability, argument pincode from {{ pincode }}, Save result as serviceability, Fail early on error ticked so the process stops if the area isn't served.
  4. Ask agent: book. Task: Offer the available slots from the serviceability check and book the one the customer picks. Wait for the user's reply ticked, Advance when this tool succeeds set to book_home_collection.
  5. Ask agent: wrap up. Task: Confirm the booking reference and remind the customer to fast for 10 to 12 hours before collection.

Result. Priya is guided through each stage in order. The serviceability check always runs before anything is booked, and the booking step only finishes when the booking tool actually succeeds.

Sending messages from a process ​

The steps inherit the agent's Action sub-nodes. To let the process send an email, SMS or WhatsApp message, for example a booking confirmation, wire the matching sender to the agent's Action port and mention it in the relevant Ask agent task.

Good practice ​

  • Give each Ask agent step one clear task with a clear end, rather than a broad instruction.
  • Use Tool steps for anything that changes records: creating a booking, taking a payment, reserving a slot.
  • Set "Advance when this tool succeeds" on every step that waits for the customer.
  • Plan for failure. Use Fail early on error or a Condition on exit_reason when a check might fail.

Next steps ​