The composable workflow editor includes an AI assistant that can build, edit, explain, and debug your workflow through a conversation. Tell it what you want — “send a Slack message when a recording about pricing comes in” — and it adds the right blocks, connects them, and asks before saving. Responses stream live for a responsive editing experience.Documentation Index
Fetch the complete documentation index at: https://docs.buildbetter.ai/llms.txt
Use this file to discover all available pages before exploring further.
The Workflow AI Assistant is available in the composable workflow editor. If you’re using the legacy template-based workflow flow, switch to the composable editor to access the chat panel.
Where to Find It
The assistant lives in a right-side Chat panel inside the workflow editor. It’s a peer of the manual editor — anything the agent does is also something you can do by hand. Edits land in memory until you click Save, exactly like manual editing. When the agent is working, you’ll see:- Animated ring around the canvas — signals the agent is busy. Try not to interrupt.
- Agent cursor over the node the agent is currently editing — makes tool-call activity visible.
What You Can Ask
The agent can:- Build a workflow from scratch — “When a recording with ‘pricing’ in the title comes in, send the summary to a Slack channel.”
- Edit existing workflows — “Change the trigger so it only fires on calls with external participants.”
- Explain a workflow — “What does this workflow do? Why is the third block needed?”
- Debug — “This isn’t firing on calls I expect. Why?”
The Agent’s Tools
Every change goes through one of these tools. Tool calls render in the chat so you can see exactly what was done.| Tool | What It Does |
|---|---|
get_workflow | Reads the current workflow definition — name, description, and graph state. Used when the agent needs context. |
list_block_types | Lists every block type in the catalog (triggers, filters, actions, etc.) so the agent picks valid names. |
add_node | Adds a new block to the canvas. Validates the type against the catalog up front; rejects unknown names with the valid list inline. |
update_node_config | Sets a config field on an existing block. Validates references and templates against the current graph. |
update_node_name | Renames a block. |
remove_node | Removes a block. Re-runs auto-layout after. |
connect_nodes | Adds an edge between two blocks. Prevents cycles, duplicates, and self-loops. |
remove_edge | Removes a connection. |
auto_layout | Re-tidies the canvas after structural changes — runs automatically after add/remove/connect. |
generate_signal_filter | Translates a natural-language description into a signal filter (same engine as the signals smart-search wand). Returns a filter envelope the workflow can use as a condition. |
Format List Block
Per the 2026-05-05 release, workflows include a Format List block for turning lists into clean, formatted text. Date formatting is configurable. The AI assistant knows about this block and will reach for it when you ask the workflow to roll up results into a readable summary.Slack Channel Picker
Slack workflow steps now let you pick a connected Slack channel directly — no manual webhook setup. The AI assistant respects this: ask for a Slack action, and it’ll set up the channel picker on the block instead of leaving a webhook field empty.Run Now & Recent Call Picker
The workflow editor lets you test a workflow without waiting for live events:- Run Now — manually trigger a workflow against a chosen recording or signal
- Recent call picker — pick from a list of recent calls to use as the test input
When the Agent Doesn’t Edit
The agent is conservative about save-able changes:- In-memory only — edits don’t persist until you click Save. The chat doesn’t bypass the save gate.
- Will not save behind your back — even after a long edit session, the agent waits for you to commit.
- Asks before destructive changes — when removing many blocks or restructuring, the agent confirms.
Long Sessions
Workflow chat is built to handle longer conversations. The system streams the agent’s response live, summarizes prior context as needed, and preserves your original request even after many turns. If the conversation gets long enough to compress, the agent keeps the original instruction intact so it doesn’t drift from your goal.Limits
- Catalog-bound — the agent can only add blocks that exist in the catalog. If you ask for a capability the catalog doesn’t expose, the agent will say so rather than improvise.
- No JavaScript execution — the agent edits the workflow graph, not the engine.
- No multi-workflow edits — each chat session is scoped to one workflow.