n8n - 💡(How to fix) Fix AgentToolV3: EngineRequest returned by sub-agent LLM tool calls is not handled, causes "not supported in Agents" error

Bug Description

When an AgentToolV3 node (sub-agent) is used as a tool for a parent agent, and that sub-agent's LLM decides to call one of its own tools (e.g. a toolWorkflow node), the execution fails with:

Error: The Tool attempted to return an engine request, which is not supported in Agents

The sub-agent's tools are never called — the error occurs earlier, when the sub-agent's LLM response is passed back through makeHandleToolInvocation in the parent agent's context.

Root Cause (traced to source)

1. Parent agent calls order-agent (AgentToolV3) via makeHandleToolInvocation
2. makeHandleToolInvocation runs nodeType.execute?.call(context) → toolsAgentExecute
3. Inside runAgent.ts, the sub-agent's LLM decides to call a tool and executes:

   // packages/@n8n/nodes-langchain/.../ToolsAgent/V3/helpers/runAgent.ts
   const actions = createEngineRequests(result.toolCalls, itemIndex, tools);
   return { actions, ... }; // returns EngineRequest (has `actions` property)

4. This EngineRequest propagates back to makeHandleToolInvocation in get-input-connection-data.ts:

   } else if (isEngineRequest(result)) {
     response = 'Error: The Tool attempted to return an engine request, which is not supported in Agents';
   }

5. The error string is returned to the parent agent as the tool result. The sub-agent's tools are never invoked.

The EngineRequest mechanism works correctly when AgentToolV3 runs as a top-level workflow node (the workflow engine handles the request/response cycle). It breaks only when AgentToolV3 is invoked from within makeHandleToolInvocation's synchronous callback, which has no way to relay EngineRequests back to the engine.

Steps to Reproduce

1. Create a workflow with:
   • A parent AI Agent node (manage-agent)
   • An AgentToolV3 node (order-agent) connected as a tool to manage-agent
   • One or more toolWorkflow nodes connected as tools to order-agent
2. Trigger the parent agent with a message that causes it to call order-agent
3. order-agent's LLM will try to call one of its own tools

Result: Error: The Tool attempted to return an engine request, which is not supported in Agents

Expected: order-agent completes its tool call loop and returns a final answer to manage-agent

Verified via Execution Data

Confirmed by inspecting SQLite execution data (execution_data table):

{ "response": "Error: The Tool attempted to return an engine request, which is not supported in Agents" }

The toolWorkflow sub-workflows are never started — isEngineRequest fires on the sub-agent's own LLM response, before any tool node runs.

Environment

• n8n version: 2.21.4 (also verified identical code in 2.22.0 — not fixed)
• Deployment: Docker (self-hosted)
• Node type: @n8n/n8n-nodes-langchain.agentTool typeVersion 3

Related

• PR #26389 proposes a fix (executeSubAgentTools / resolveSubAgentEngineRequests) but has not been merged as of 2.22.0
• The issue exists for any tool type connected to AgentToolV3 that triggers an LLM tool call internally (not just MCP/Vector Store/HITL as listed in that PR)

Workaround

Convert each AgentToolV3 sub-agent into a standalone n8n workflow called via a toolWorkflow node. Inside the standalone workflow, a regular AI Agent node handles EngineRequests correctly via the workflow engine's request/response cycle.