Skip to content

Workforce ​

What it is ​

A Workforce is a team of Agents, Tools and logic on a canvas: who does which part, who passes work to whom, and where a person approves. Use it when one Agent is not enough: several specialist Agents pass work to each other, a chain of Tools must always run after a step, or a person must approve between steps. Find it in the left menu under Build → Workforces.

Start with one Agent with Tools and Skills; the create dialog asks "Could one agent do it?" because one Agent is usually faster and cheaper.

Key concepts ​

Nodes (from the Add node bar)

NodeWhat it does
StartThe entry, one per Workforce ("When new work arrives"). Holds the Input values a run receives and the Handoff limit.
AgentA published chat Agent of the Workspace. Agents are shared: editing one here edits it everywhere it is used.
ToolA published Workspace Tool, or an App action, that always runs at that point.
ConditionA gate: when it is true the next steps run, when it is false that branch stops. By LLM: one sentence the LLM answers from the task data. By rules: rules on earlier outputs (equals, is not, greater than, less than, contains), matching All or Any.
JoinWaits for all incoming parallel branches, then continues with their outputs together.
NoteText for people; never runs.

Connections — what happens after a step. Pick the type when you connect two nodes:

TypeMeaningBetween
Next stepThe target always runs after the source finishes. Nothing comes back to the source.Start, Agent, Tool, Condition, Join → Agent, Tool, Condition, Join
DelegateThe source Agent calls the target Agent when needed and gets the result back, like a Tool. The source keeps replying to the sender.Agent → Agent
HandoffThe source Agent hands the conversation to the target, who replies to the sender from then on.Agent → Agent
  • Delegate and Handoff need When to call / When to hand off: plain words the source Agent reads to decide. Write them like briefing a colleague.
  • A delegated Agent sees Parameters only (default) or With conversation. A Handoff passes the Full conversation history (default) or a Summary.
  • Two or more Next step connections out of one node run in Parallel. A step that several parallel branches lead into needs a Join before it.

Conversation — the Coordinate, Handoffs and Route templates start with a conversation (a Conversation chip on Start): the Workforce talks with a sender. At least one Agent must then reply directly (marked Replies directly): one reached by Next step from Start or from another Agent, or the target of a Handoff. An Agent reached only by Delegate, or by Next step from a Tool, Condition or Join, works in the background. Workflow and Blank canvas have no sender: a run starts from Start with its inputs.

Map parameters — a connection into a Tool or an Agent fills the target's inputs. Type {{ to insert a value: the latest message, an earlier Agent's reply or Tool output, a Start input ({{input.<name>}}), contact fields, {{today}} or {{now}}. An empty Delegate parameter is filled by the calling Agent.

Approval — only on a Next step connection into a Tool:

  • Auto run — no approval needed.
  • Needs approval — drafts the call, then waits for an Approver (a Workspace member who can operate Agents; when Not chosen, the Workspace default recipient or, if none is set, the person who created the Automation). Approvals are answered in the Agent Inbox.

A new connection into a write Tool starts on Needs approval, into a read Tool on Auto run. Removing approval from a write Tool asks you to confirm and is written to the Workspace audit log.

Auto-run limit — on a connection, how many times it may run in one task. Into a Tool, past the limit it asks for approval; on a Delegate, past the limit the source can no longer call the target. A connection back to an earlier node needs this limit (a Handoff back is bounded by the Handoff limit instead).

How it works ​

Status

StatusMeaning
DraftNever published.
PublishedHas a production version.
ArchivedRead-only.

Draft, test, publish

  1. Edits on the canvas are local until you press Save (⌘S); the header shows "(Unsaved)". Save stores them as the draft.
  2. The canvas lists issues to fix. Publish is blocked until there are none: every Agent and Tool slot holds a real, published Agent or Tool; every node is reachable from Start; the Apps its Tools use are connected; loops have an Auto-run limit; parallel branches meet in a Join; rules use fields that exist; approvers are members; a conversation has an Agent that replies directly. "A required input is not mapped" is a warning only.
  3. Test runs the saved draft (see below).
  4. Publish makes the draft the production version. A Workforce always runs the published version of each of its Agents and Tools.

Until the first publish, a Getting started card shows the steps: pick an Agent or Tool for every slot, write "when to call" on Agent connections, connect every node from Start, fix every issue, run one test, publish.

Test runs

Open Test in the header. Fill in the Inputs for this run, or write a message as the sender would, and press Run. Nodes and connections light up on the canvas as the run goes through them.

  • Safe test run (on by default): write Tools are simulated ("Safe, writes simulated"); reads are real.
  • With it off, writes really happen ("Live writes").
  • An approval in a test run is answered by you in the test panel.
  • Run statuses: Running, Waiting for you, Done, Failed, Handed to a person. Each step shows Running, Waiting, Done, Failed or Rejected, and a failed step says why (for example "The Tool failed." or "A required parameter of this step is missing.").
  • A test run needs the same Agents and Tools to be published, but not an Automation.

Production runs

A published Workforce runs when an App event Automation targets it (set up on the Automations page), for example a new Haravan order (each App's events are listed on its page under Apps). Message channels are answered by an Agent, not a Workforce. Each run's approvals wait in the Agent Inbox; the Workforces list shows how many are Waiting and links to Open Approvals.

A run stops when it reaches its step limit or time limit. Past the Handoff limit (default 5 handoffs in one turn) the work goes to a person.

Common tasks ​

Create a Workforce

  1. Go to Workforces and press New workforce.
  2. Enter a Name and, optionally, Instructions (the goal and shared rules; saved as the Workforce's description).
  3. Pick a template: Blank canvas, Coordinate (one Agent delegates to specialists), Handoffs (one Agent per stage), Route (an LLM Condition sends the request to the right Agents, then a Join), or Workflow (steps always run in the order drawn).
  4. Press Create workforce. Templates place placeholder nodes: click each to pick a real Agent or Tool.

The Ask FLAE Assistant tab of this dialog (also opened by Create with Assistant) shows a sample proposal only. To have the Assistant build a Workforce for you, ask it in Chat (see Chat with the Assistant); it saves a draft for you to review, pick approvers and publish.

Build the graph

  1. Drag or click Agent, Tool, Condition, Join or Note from the Add node bar. Create new agent / Create new tool open the builders.
  2. Connect two nodes, choose Next step, Delegate or Handoff, and press Connect.
  3. Click a connection to set its parameters, approval, approver and Auto-run limit.
  4. Click an Agent node to read its setup; you can change its instructions from the canvas. If the Agent has no draft, Save changes saves and publishes it at once ("Agent saved and published."); if it has a draft, the change goes into that draft for you to publish in the Agent builder.

Limits and rules ​

  • Who can do what: Workspace Admins and Builders (and Organization Owners and Admins) create, edit, test and publish Workforces. Other members can view them. Nothing can be changed in an archived Workspace.
  • Only published chat Agents can be placed; voice Agents cannot. Draft Agents and Tools show "Not published" and block Publish and test runs.
  • Start can hold up to 20 inputs. The Handoff limit is 1 to 50 per turn.
  • A Workforce with a conversation cannot have a Join after a Needs approval connection.
  • The Assistant can create and change draft Workforces but cannot publish them.
  • The Workforce page has no archive, delete or version history controls yet.
  • Agents — the Agents you place on the canvas.
  • Tools — fixed steps and their approval.
  • Knowledge — Internal Knowledge Bases are read in Workforce steps without a conversation.
  • Automations — start production runs from App events.
  • Agent Inbox — where approvals of Workforce runs are answered.
  • Apps — App actions a Tool node can run.