Mastra•2mo ago

HELP: sub agent calls supervisor agent tools

i have a multi agent setup using subagents wrapped as tools. The reason why i have pivoted to such a setup is mainly because of

agent.network()

agent.network()

agent.network()

agent.network() fails to wait for user confirmation at routing agent level for a suspended workflow initiated by the subagent.

So i am following the supervisory agent method, Everything works well with this setup as you can see in the brief video the traces are perfect as expected

supervisor

supervisor

supervisor

supervisor ->

outcome agent

outcome agent

outcome agent

outcome agent ->

workflow

workflow

workflow

workflow including feedback, however sometimes (could not identify when but seems like mostly at HITL steps) sub agent calls the wrapped tool to call itself, and hence fails by

Error: Tool delegateToOutcomeAgent not found

Error: Tool delegateToOutcomeAgent not found

Error: Tool delegateToOutcomeAgent not found

Error: Tool delegateToOutcomeAgent not found

This is my supervisor agent

const chatAgent = new Agent({
  name: "Chat Agent",
  memory: async ({
    runtimeContext,
  }: {
    runtimeContext: RuntimeContext<Context>;
  }) => {
    return await getMemory(runtimeContext.get("schemaId"));
  },
  model: getClaudeModel("claude-3-7-sonnet-latest"),
  tools: {
    delegateToOutcomeAgent,
  },
  instructions: chatAgentInstruction,
});
export { chatAgent };

const chatAgent = new Agent({
  name: "Chat Agent",
  memory: async ({
    runtimeContext,
  }: {
    runtimeContext: RuntimeContext<Context>;
  }) => {
    return await getMemory(runtimeContext.get("schemaId"));
  },
  model: getClaudeModel("claude-3-7-sonnet-latest"),
  tools: {
    delegateToOutcomeAgent,
  },
  instructions: chatAgentInstruction,
});
export { chatAgent };

const chatAgent = new Agent({
  name: "Chat Agent",
  memory: async ({
    runtimeContext,
  }: {
    runtimeContext: RuntimeContext<Context>;
  }) => {
    return await getMemory(runtimeContext.get("schemaId"));
  },
  model: getClaudeModel("claude-3-7-sonnet-latest"),
  tools: {
    delegateToOutcomeAgent,
  },
  instructions: chatAgentInstruction,
});
export { chatAgent };

const chatAgent = new Agent({
  name: "Chat Agent",
  memory: async ({
    runtimeContext,
  }: {
    runtimeContext: RuntimeContext<Context>;
  }) => {
    return await getMemory(runtimeContext.get("schemaId"));
  },
  model: getClaudeModel("claude-3-7-sonnet-latest"),
  tools: {
    delegateToOutcomeAgent,
  },
  instructions: chatAgentInstruction,
});
export { chatAgent };

sub agent:

export const outcomeAgent = new Agent({
  name: "Outcome Agent",
  description: `This agent helps users create clear, measurable business outcomes through guided conversation.
    It gathers context, evaluates information quality, and generates concise, actionable outcomes
    with specific timelines and expected results. Use this agent when users want to define goals,
    objectives, or measurable business targets.`,
  memory: async ({
    runtimeContext,
  }: {
    runtimeContext: RuntimeContext<Context>;
  }) => {
    return await getMemory(runtimeContext.get("schemaId"));
  },
  model: getClaudeModel("claude-3-7-sonnet-latest"),
  instructions: outcomeAgentInstruction,
  tools: {
    startOutcomeWorkflow: startOutcomeWorkflowTool,
    resumeOutcomeWorkflow: resumeOutcomeWorkflowTool,
  },
});

export const outcomeAgent = new Agent({
  name: "Outcome Agent",
  description: `This agent helps users create clear, measurable business outcomes through guided conversation.
    It gathers context, evaluates information quality, and generates concise, actionable outcomes
    with specific timelines and expected results. Use this agent when users want to define goals,
    objectives, or measurable business targets.`,
  memory: async ({
    runtimeContext,
  }: {
    runtimeContext: RuntimeContext<Context>;
  }) => {
    return await getMemory(runtimeContext.get("schemaId"));
  },
  model: getClaudeModel("claude-3-7-sonnet-latest"),
  instructions: outcomeAgentInstruction,
  tools: {
    startOutcomeWorkflow: startOutcomeWorkflowTool,
    resumeOutcomeWorkflow: resumeOutcomeWorkflowTool,
  },
});

export const outcomeAgent = new Agent({
  name: "Outcome Agent",
  description: `This agent helps users create clear, measurable business outcomes through guided conversation.
    It gathers context, evaluates information quality, and generates concise, actionable outcomes
    with specific timelines and expected results. Use this agent when users want to define goals,
    objectives, or measurable business targets.`,
  memory: async ({
    runtimeContext,
  }: {
    runtimeContext: RuntimeContext<Context>;
  }) => {
    return await getMemory(runtimeContext.get("schemaId"));
  },
  model: getClaudeModel("claude-3-7-sonnet-latest"),
  instructions: outcomeAgentInstruction,
  tools: {
    startOutcomeWorkflow: startOutcomeWorkflowTool,
    resumeOutcomeWorkflow: resumeOutcomeWorkflowTool,
  },
});

export const outcomeAgent = new Agent({
  name: "Outcome Agent",
  description: `This agent helps users create clear, measurable business outcomes through guided conversation.
    It gathers context, evaluates information quality, and generates concise, actionable outcomes
    with specific timelines and expected results. Use this agent when users want to define goals,
    objectives, or measurable business targets.`,
  memory: async ({
    runtimeContext,
  }: {
    runtimeContext: RuntimeContext<Context>;
  }) => {
    return await getMemory(runtimeContext.get("schemaId"));
  },
  model: getClaudeModel("claude-3-7-sonnet-latest"),
  instructions: outcomeAgentInstruction,
  tools: {
    startOutcomeWorkflow: startOutcomeWorkflowTool,
    resumeOutcomeWorkflow: resumeOutcomeWorkflowTool,
  },
});

Aman ZishanOP•11/11/25, 2:20 PM

use model: claude-3-7-sonnet-latestclaude-3-7-sonnet-latest
mastra versions:
"@mastra/core": "^0.24.0",
"@mastra/evals": "^0.14.3",
"@mastra/libsql": "^0.16.2",
"@mastra/loggers": "^0.10.19",
"@mastra/memory": "^0.15.11",

Aman ZishanOP•11/11/25, 2:21 PM

@roamin friendly nudge on this

Mastra Triager•11/12/25, 11:46 AM

Created GitHub issue: https://github.com/mastra-ai/mastra/issues/10007

GitHub

HELP: sub agent calls supervisor agent tools · Issue #10007 · mas...

This issue was created from Discord post 1437807991876354271: i have a multi agent setup using subagents wrapped as tools. The reason why i have pivoted to such a setup is mainly because of agent.n...

Aman ZishanOP•11/12/25, 1:04 PM

I think additionally since the memory is being shared ; the sub agent call within the wrapped tool actually creates a message. This causes duplication when messages are persisted and fetched on refresh as sub agent writes to the thread on agent.generate()agent.generate()

is there a solution to that?
The one i stumbled upon was to remove memory from sub agent but that does not work as sub agent need to access the memory to be able to have the shared context for the HITL flows

_roamin_•11/12/25, 2:55 PM

Hi @Aman Zishan ! Unfortunately, tool calls always have a chance to fail because it's up to the LLM to provide the correct arguments, however, you can help the LLM by guiding it in the instructions. You could tell that it CANNOT call the agent with id blablabla (or other instructions).

_roamin_•11/12/25, 2:57 PM

For you second question, I need to check, you might need to pull the messages from memory manually and add them to your messages

Aman ZishanOP•11/12/25, 2:57 PM

I did try that but it seems to fail even then, the thing is i have not mentioned delegateToOutcomeAgentdelegateToOutcomeAgent in the sub agent agent so it should not have access to that tool any way right?

__roamin_Hi @Aman Zishan ! Unfortunately, tool calls always have a chance to fail because...

Aman ZishanOP•11/12/25, 3:00 PM

I think the problem here is the sub agent calling a tool which was not listed under that particular agent, that is different to what you explained right

__roamin_For you second question, I need to check, you might need to pull the messages fr...

Aman ZishanOP•11/12/25, 3:03 PM

That is what i am doing right now; the problem is that since two agents have access to the same thread the sub agent also writes the message to that thread hence the order gets messed up. similar issue: help-bugs-problemsCan you pass threadId/resourceId memory between multiple agents? or workflow runs?

It would be good if you could guide me to some example on setting up a reliable multi agent sustem with HITL involved cause i have tried different methods and this one is by far the most reliable way to accomplish such a setup.

2 problems left are

Sub agent calls supervisor agent tool even though sub agent does not have the tool listed
Duplicate messages entry created by subagent (i want the subagent to share the memory and use agent.generate() but should not write to the thread)

HELP: sub agent calls supervisor agent tools

Similar Threads

Similar Threads

Similar Threads