Skip to main content

#MuleSoft For Agentforce3 discussing

Questions and answers about MuleSoft for Agentforce, best practices, and use cases.

I deployed an Agent Network 2.0 project (one AgentScript broker, OpenAI LLM, MCP tools) to a managed Omni Gateway in the CloudHub 2.0 shared space. I’m on a free trial org and used agent-network project deploy -g <gateway>.

The deploy succeeds. The app shows APPLIED, RUNNING, 1/1, and the broker compiles and mounts at /restaurant_broker. The pod logs this warning at startup:

LLM provider 'rm_openai_llm' at 'http://testgateway:8082/…/rmOpenaiConnection': LLM provider responded with HTTP 400 on /responses. Server is reachable but the request was rejected -- skipping strict validation.

Every request to the broker URL (agent card, root, A2A SendMessage) returns an empty HTTP 503 from “Anypoint Flex Gateway”. The pod logs show no incoming requests.

Still 503 after:

  • a clean redeploy
  • a fresh MCP server
  • --disable-tracing
  • removing the Tracing policy
  • logging.level ERROR

The same project works locally in Docker. The agent card returns 200 there.

Has anyone run a broker on a free trial? Is there a shared-space limit, or a way to see why the gateway returns 503? 

 

Instead of shared space, I tried creating a private space. But I could not create a network in a private space in the trial account.  

 

We have a community demo coming up soon and we need to make it working soon.  

 

Any suggestion / help is appreciated.  

 

#MuleSoft For Agentforce  #Agentforce

Using Anypoint Code Builder with agent-fabric-transformation 1.6.0 (agent-network.yaml, agentNetwork: 2.0.0).

Every Deploy logs this warning for my AgentScript broker:

Warning: Broker 'customer_support_broker' does not declare an 'authorization' configuration, so it will fall back to the 'user-context-propagation' policy defaults: userIdExpression '#authentication.clientId' and allowAnonymousAccess true (requests whose principal cannot be determined are forwarded anonymously). To change this, set authorization under the broker's interfaces.a2a section, e.g.:

brokers: 

    customer_support_broker: 

       interfaces: 

          a2a: 

             authorization: 

                 principal: '

#attributes.clientId

' 

                 allowAnonymousAccess: false 

 

At runtime this also throws a recurring "User ID expression resolved to invalid non-string value" warning, since I have no Client ID Enforcement policy configured (I don't need real caller identity — this broker isn't meant to authenticate clients).

Problem: following the warning's own suggested fix fails Build validation:

Message: brokers/customer_support_broker/interfaces/a2a must NOT have additional properties

So authorization isn't actually a field the current agent-network.yaml schema (v2, as validated by agent-fabric-transformation 1.6.0) accepts under interfaces.a2a — it only allows card and policies.

I then tried explicitly declaring the policy under policies.inbound instead:

brokers: 

     customer_support_broker: 

         interfaces: 

             a2a: 

                 policies: 

                     inbound: 

                         - policy: 

                             ref: 

                                 name: user-context-propagation 

                                 namespace: 68ef9520-24e9-4cf2-b2f5-620025690913 

                             configuration: 

                                 userIdExpression: '

#"anonymous"

' 

                                 allowAnonymousAccess: true

(namespace taken from the groupId on the built connection.json's policyRef for this same policy). This passes schema validation but fails with:

Cannot resolve reference user-context-propagation of kind policy (http://a.ml/vocabularies/data#invalid-reference-like)

Questions:

  1. What's the actual supported syntax to override this auto-injected default policy's configuration from agent-network.yaml?
  2. Is the authorization block from the warning message a planned/upcoming field not yet implemented in 1.6.0, or am I missing something?
  3. Alternatively, is there a supported way to just suppress the userIdExpression warning without adding real Client ID Enforcement?

Happy to share the full agent-network.yaml if useful. 

 

#MuleSoft For Agentforce  #MuleSoft  #MuleSoft Anypoint Code Builder

1 answer
  1. Oct 8, 8:43 PM

    @Miguel Ribeiro the warning and the schema are describing two different shapes, and 1.6.0 is checking the schema.

    Agent Network 2.0 documents interfaces.a2a as card plus policies. There is no authorization property on that object. That's why the example in the warning dies with must NOT have additional properties. I wouldn't treat that example as a field this build forgot to mention. The reference I can stand on only lists card and policies.

    A policy gets onto the broker after you declare it, then point at it with a binding reference, ref.name. Or you put an inline binding under policies.inbound with policy.ref and policy.configuration. Your second try is that inline shape, and it got past the schema. Cannot resolve reference user-context-propagation of kind policy means that name isn't a policy declared in this yaml. Copying the namespace off the generated connection.json doesn't turn the auto-injected policy into a reference you can override. The injector still owns the default.

    The runtime warning is the default expression reading authentication.clientId, which is empty when you don't have Client ID enforcement. That value isn't a string, so you get invalid non-string value. If no ingress auth policy is configured, the broker still orchestrates. What you lose is task management on tasks created with no userId. I don't have a 1.6.0 switch that silences that expression and keeps the broker anonymous. A constant user id would have to live on a policy reference this file can actually resolve.