Workflows
Automated flows built on a canvas, for when a task has fixed steps instead of a conversation.
A Workflow is a sequence of steps built on a canvas, connecting nodes that receive data, process it and return a result.
Where an Agent converses, a Workflow executes. The person using it fills in an input form and gets an output, with no dialogue in between.
Agent or Workflow
The question that separates the two is whether the path is known in advance:
- Agent: the person asks open questions and the model decides what to do on each message. Good for questions, support and research.
- Workflow: you know the steps and their order. Good for processing a document, extracting fields, querying a system and generating a file.
A Workflow is more predictable and easier to audit. An Agent covers cases you didn't anticipate.
The nodes
The canvas offers nodes grouped by purpose:
- Input and Output: the two ends of the flow. Input defines the form fields the person fills in; Output defines what they get back.
- LLM: calls an AI model with a prompt built from what came before.
- Documents: File extractor, CPF extractor, Vector database and File generator.
- Logic: Condition, Iteration and Aggregation, for branching, repeating over lists and choosing between results.
- Integration: HTTP request, Send email and Integration, for talking to outside systems.
- Code: Code executor, for transformations no ready-made node covers.
The Vector database node is what gives a Workflow access to Collections, and it requires at least one configured Collection.
Draft, publishing and runs
Like Agents, Workflows have a draft and a published version. Only the published one is available for other people to run.
Each execution becomes a run with its own record: input, output, steps taken and errors. That is where you investigate a strange result, step by step, instead of guessing.
Where people run them
Published Workflows show up in the Agents area of the Chat module, under the icon right below Chat in the area bar.
Whoever opens a workflow there sees the form defined by the Input node, runs it and follows the result. No Admin access is needed for this.
The area's name is confusing: Agents, in Chat, lists published Workflows, not conversational Agents. Conversational Agents are picked from the selector inside the Chat screen itself.
Next steps
- Understand the Collections the vector node queries in Collections and knowledge.
- See the tools the integration nodes reach in MCPs and Integrations.