Workflow Tool: Call Another Workflow from an Agent
The Workflow Tool node turns another workflow into one tool an agent can call. Attach it to an Agent or Supervisor Tool port and pick the workflow. The child runs in the same run, in this process, with the agent's call as its input; its output is the tool result.
Concepts: Workflows vs agents explains why a fixed workflow behind an agent is a common design.
Use it when a step is already a workflow (a lookup with retrieval and a content rule, a ticket triage flow) and you want an agent to decide when to run it.
What the agent sees
The tool is derived from the child workflow's entry trigger:
- An Expose as MCP Tool trigger: its tool name, description and input schema.
- Otherwise a Sub-Workflow Trigger: its description and input schema; the name comes from the workflow name.
- Otherwise one free-text
inputargument, which the child also gets as its user prompt ().
Tool name and Description for the model on the node override the derived ones. The panel previews the result. Each tool on an agent needs its own name: two Workflow Tools that end up with the same name (the same workflow twice, or the same Tool name) fail the agent node, and validation reports duplicate_agent_tool_name when it can tell from the workflow.
The child gets the arguments as its input (each argument at the top level, and all of them under arguments).
Limits
Workflow tools follow the same rules as Execute Workflow:
- Depth and cycles. At most 10 workflows deep; a workflow already on the call stack (the agent's own workflow included) is refused. Both come back to the agent as a tool error.
- Project scope. The child must be in the same project, or shared to it with execute rights. It runs with its own project's secrets, knowledge bases and session stores. A workflow the run may not use fails the agent node with the reason.
- Time. A call is bounded by the run's own timeout. Set Timeout per call on the node for a shorter limit: the call then fails with a message the agent can act on, and the child is told to stop.
- Approvals. If the child stops for an approval, the call fails with a message the agent can act on; the child is not kept waiting. Put the approval in its own run, or approve the call in the parent instead (below).
Governance
A workflow tool is governed as the MCP server workflow:<workflowId>:
- Tool rules naming server
workflow:<id>deny it, ask for approval, or allow it, like any MCP tool. A denied workflow tool is hidden from the model. - Write-tool approval. A workflow tool counts as a write unless the child declares itself read-only: Read-only workflow on its Sub-Workflow Trigger or Expose as MCP Tool trigger, set by the child's own editors. A caller cannot mark it. So a workflow with write-tool approval on asks before each call to a child that has not declared it.
- Write controls reach the child. The calling workflow's write-tool approval and its No write tools switch also apply inside the child's run, on top of the child's own settings; they only ever add checks. They also reach every workflow further down: a Workflow Tool inside the child, and an Execute Workflow node in the child or below it. A child declared read-only that still calls a write tool, at any depth, therefore meets the caller's approval there, and the call fails with a message saying it needs one.
- Content rules check the call's arguments and result like any tool call, under the parent run's rule sets; the child runs under them too.
- Labels. The labels of the agent's context at the call go to the child as its input labels, and the child's output labels come back on the result (on a failed run, the labels of what its failed step read, since the error can quote it). Untrusted content in the call therefore reaches the child's own tool rules.
- Telemetry and replay. Each call writes an MCP tool call row (adapter
l2m-workflow). A recorded run's cassette holds the call, so a replay answers it from the recording and never runs the child. - Spend. The child's model calls count in the run's usage and budgets as soon as the call ends, each priced by its own model, also when the agent then fails or is stopped (see Budgets: sub-workflows). Before each call the run's block budgets are checked, counting the calls before it; a used-up one refuses the call with a tool error. A child that is still running when its call times out is counted when it ends.
Across instances
To call a workflow on another L2M instance, or to give it its own execution, budget and applied rule sets, expose it through a toolset on the MCP gateway and add that endpoint as an HTTP MCP server. See Expose a Workflow as an MCP Tool.