Contact Us 1-800-596-4880
  1. Homepage
  2. Building Agent Networks (2.0)

  3. Troubleshoot Agent Networks

Troubleshoot Agent Networks

Resolve common agent network errors and connectivity issues by enabling verbose logging, message logging policies, and distributed tracing. Using logs and distributed traces, you can observe agents and MCP servers and see how requests flow through brokers, which tools and policies apply, and where failures or latency originate.

Use Agent Visualizer to view metrics, logs, and traces for agents and MCP servers in your network.

Enable Logging to Debug Agent Brokers

Enable logging and log levels from CloudHub 2.0 or Runtime Fabric.

Agent brokers provide two levels of logging detail:

  • INFO: Logs node transitions, MCP tool discovery, agent card retrieval, and related operational events. INFO-level logs don’t use the INSECURE-LOGGING level.

  • DEBUG: Logs LLM reasoning, LLM input and output, tools invoked, and all external calls made by the broker. All DEBUG-level logs use the INSECURE-LOGGING level.

To enable logging:

  1. In Runtime Manager, navigate to the application for the agent broker.

  2. Choose one of these options to configure logging:

  3. To view operational logs, set the log level to INFO.

  4. To view detailed LLM and tool execution logs, set the log level for INSECURE-LOGGING to DEBUG.

INSECURE-LOGGING logs content that can contain sensitive information, such as LLM reasoning, and inputs and outputs from agents and assets defined in your agent network. This content is very likely to contain personally identifiable information (PII). Disable INSECURE-LOGGING after you finish troubleshooting.

Enable Message Logging Policy

Edit agent-network.yaml to set message logging for MCP servers and agents.

Message logging uses the Omni Gateway Message Logging policy. In the connection section of the agent or MCP server instance you want to log, add this logging policy inside policies:

policies: ## add the following inside the policy setting
    - ref:
        name: message-logging
        namespace: business_group_UUID
      configuration:
        loggingConfiguration:
          - itemName: "Payload"
            itemData:
              message: "#[payload]"
              firstSection: true
              secondSection: true
              level: "INFO"
          - itemName: "Headers"
            itemData:
              message: "#[attributes]"
              firstSection: true
              secondSection: true
              level: "INFO"
On non-US control planes, the value for the namespace parameter can change. In Exchange, check the business group in the API Gateway Message Logging Policy Template.

Troubleshooting Common Issues and Errors

Use the steps in this section to diagnose and fix common problems in agent networks.

Issue: Traces Not Shown

If Anypoint Monitoring isn’t showing traces for agents and MCP servers, check these settings:

  • In API Manager, verify you’ve enabled the tracing policy for all brokers and agents.

  • In the managed Omni Gateways, verify that you’ve enabled Distributed Tracing for ingress and egress.

Issue: No Agent Network Commands in Anypoint Code Builder

If you installed Anypoint Code Builder successfully, but you don’t see agent network commands in the interface, verify that you have the latest version of the Anypoint Extension Pack. Then, restart VS Code.

Issue: Linting Errors in Agent Script File

If you see linting errors in your .agent file in VS Code, you likely have the Agentforce DX extension pack installed, and it’s conflicting with Anypoint Code Builder. This extension pack includes a dependency on the Agent Script extension that conflicts with the Agent Script extension built into Anypoint Code Builder.

To resolve this conflict, uninstall the Agentforce DX extension pack from VS Code and restart the IDE. Anypoint Code Builder already includes all Agent Script functionality, so you don’t need a separate extension.

Issue: No Agent Network Deployment Targets

If you attempt to deploy an agent network, but you don’t see any deployment targets in Anypoint Code Builder or Anypoint CLI, then you need to set up your target deployment space and set up gateways. For instructions, see Get Started with Agent Networks.

Issue: Changes Don’t Take Effect After Redeployment

If you update agent-network.yaml or any .agent files but changes aren’t reflected after you redeploy, the deployment used the last compiled target/ directory. Every source change requires a full rebuild.

After any source file change, run the full pipeline:

  1. Build

  2. Publish

  3. Deploy

Error: "Asset belongs to another project" (Error 2006)

If the publish command fails with this error (code 2006), an asset in your project has the same name as an asset that belongs to a different project in the same organization. Asset names must be unique across all projects in an org, and you can’t reuse an asset name that another project already owns.

To resolve this conflict:

  • In agent-network.yaml, rename the conflicting asset to a unique name.

  • Check the other project referenced in the error to verify which assets it owns. If you own the conflicting project, remove or rename the asset in that project before you republish.

Error: CrashLoopBackOff or Exit 139 with No Clear Error

If your deployment keeps restarting without a clear error, the broker is most likely crashing on a bad upstream connection before your agent network configuration is parsed. Common causes include:

  • An expired API key

  • An incorrect MCP server URL

  • An MCP server returning JSON instead of SSE

  • A missing A2A agent card

Before you deploy, validate every connection with live probes. Don’t troubleshoot your agent network configuration until you confirm all upstream connections are healthy.

Issue: Broker Cannot Reach Downstream Agents (Last-Mile Security Enabled)

If your broker deploys successfully and answers general questions, but fails whenever it routes a request to a downstream agent or MCP server, check if Last-Mile Security is enabled on the Omni Gateway used for outbound (egress) traffic.

Agent networks route outbound traffic directly to the pod through the egress gateway (port 8082), which bypasses the ingress-to-pod hop that Last-Mile Security protects. As a result, agent networks are not compatible with Last-Mile Security on the egress gateway.

To resolve:

  • In Runtime Manager, go to Omni Gateways > <the gateway in use> > Settings, and disable Last-Mile Security.

  • If Last-Mile Security is required for other applications on that gateway, deploy your agent network to a separate egress gateway that has Last-Mile Security disabled.

This is a behavior change from Agent Network version 1.0, which routed outbound traffic through the CloudHub ingress gateway (port 8081) and was therefore unaffected by Last-Mile Security. A gateway that worked under version 1.0 with Last-Mile Security enabled will fail under version 2.0.

Error: "Cannot complete task due to issue accessing reasoning engine"

This is a generic error that the broker issues when the LLM fails. To troubleshoot this error, enable the message logging policy on the LLM instance, and then review the logs to determine the cause of the error.

Issue: LLM Responds Without Calling the Agent or Tool

If the broker responds with information from the LLM’s training data instead of the LLM delegating to your agent or calling an MCP tool, the LLM’s instructions aren’t explicit enough about delegation. This is especially common with Gemini on widely-known topics.

In the broker’s system instructions:

  • Add a hard directive, such as MUST delegate IMMEDIATELY and NEVER answer yourself.

  • Map every intent to a specific named action.

Error: Runtime System Limit Exceeded

If your agent network encounters errors related to exceeding iteration limits, tool call limits, or state size limits, you can configure runtime system limits in your exchange.json file. These variables control graph execution behavior and help prevent infinite loops or excessive resource usage.

Common limit-related errors include:

  • Maximum handoff iterations exceeded

  • Maximum reasoning iterations exceeded

  • Maximum tool call count exceeded

  • State size limit exceeded

To adjust these limits, add runtime system limit variables to the metadata.variables section of your exchange.json file. For detailed information about configuring these limits, see Runtime System Limits Configuration.

Error: "HTTP error 404: Agent card not found at /.well-known/agent-card.json"

You can encounter this error in two different scenarios:

  1. The agent card URL is incorrect. In this case:

  2. The A2A agent doesn’t use A2A protocol version 1.0.0. In this case:

    • If you created the agent with MuleSoft, verify that you are using the latest version of the Anypoint Connector for A2A (A2A Connector).

    • If you are using an external agent, verify that it uses the A2A protocol version 1.0.0.

Error: "401 Unauthorized" When a Broker Calls an Agent

If the broker returns 401 Unauthorized error when it routes a request to a downstream A2A agent, the A2A protocol version declared for that agent in agent-network.yaml might not match the agent’s actual runtime protocol version.

When the declared version doesn’t match the real version, the platform applies the wrong A2A Agent Card policy. That policy doesn’t rewrite the agent card’s url field to the egress proxy URL, so the A2A SDK posts send_message directly to the external gateway URL and bypasses the egress proxy. Because the egress proxy is what injects the client_id and client_secret headers required by the Client-ID-Enforcement policy, the request arrives without credentials and is rejected with a 401 error.

To resolve this error:

  1. Confirm the agent’s real A2A protocol version. Request the agent card in a browser or Postman by appending .well-known/agent-card.json to the connection URL, and check the protocolVersion field.

  2. Update the connection definition in agent-network.yaml so the declared protocol version matches the agent’s real version.

  3. Rebuild, republish, and redeploy the agent network.

A protocol version mismatch surfaces as an authentication (401) error, not as a version error. Verify the protocol version whenever a downstream agent call fails with 401 even if you’ve correctly configured the the egress auth policy.

See Also