Key takeaways
- Salesforce launched MuleSoft Agent Fabric on 25 September 2025 to discover, orchestrate, govern and observe AI agents regardless of where they are built; its first release followed in October 2025.
- Its parts are Agent Registry with scanners, MCP Bridge and A2A connectors, Omni Gateway policies with a model proxy for LLM cost control, Agent Broker networks defined in Agent Script, and Agent Visualizer.
- MuleSoft publishes no price for Agent Fabric and asks customers to contact their account team; agent networks run on CloudHub 2.0, Runtime Fabric or, since November 2026, Hyperforce.
What MuleSoft Agent Fabric is
MuleSoft calls Agent Fabric one control plane to see, connect, govern and optimize AI across an enterprise, whether agents come from Salesforce, Amazon, Google, Microsoft or your own teams. Salesforce announced it on 25 September 2025, MuleSoft's release notes record the first release on 3 October 2025, and MuleSoft says new capabilities roll out monthly.
| Part | What it does |
|---|---|
| Agent Registry | One inventory of agents, MCP servers and APIs, filled by scanners on platforms such as Amazon Bedrock, Google Vertex AI, Microsoft, Databricks, ServiceNow and Kong. |
| MCP Bridge and A2A | Existing APIs and SaaS apps become MCP servers without custom code, and A2A Bridge makes agents such as Agentforce agents A2A-compliant. |
| Omni Gateway | One enforcement point for authentication, access and security rules on agents, MCP servers, APIs and LLM traffic. |
| Model proxy | Spend tracking, budgets, caps and model wallets, with semantic routing, caching and payload trimming. |
| Agent Broker | Routes tasks to the right agents and tools in an agent network. |
| Agent Visualizer | Maps agents, brokers, MCP servers and LLMs, with live flows, latency, errors and alerts. |
How agent networks are built
- An agent network is a coordinated group of agents, brokers, LLMs and MCP servers, defined in a YAML project with a registry of the assets, connections and policies it uses.
- Since July 2026 brokers are defined in Agent Script (.agent files) as graphs, with deterministic nodes for fixed paths and probabilistic nodes where an LLM decides, and agents talk over version 1.0 of the A2A protocol.
- Teams design and test networks in Anypoint Code Builder or MuleSoft Vibes, and deploy them with the Anypoint CLI to CloudHub 2.0, Runtime Fabric or Hyperforce.
How it governs agents and cost
- The same authentication, access and security rules apply to every agent, MCP server and LLM call, defined once and applied across gateways.
- Security features include rogue agent detection and containment, a kill switch, external vaults (AWS Secrets Manager, Azure Key Vault and HashiCorp Vault) and Akamai risk correlation.
- LLM spend is tracked against budgets, caps and model wallets; MuleSoft notes that budgets are approximate, as usage can briefly pass a limit before requests are blocked.
How it is bought
MuleSoft publishes no price for Agent Fabric; its product page asks customers to get in touch with their account team. MuleSoft describes Agent Fabric as enhancements to existing MuleSoft products plus new capabilities built on MuleSoft, so start the conversation from your current MuleSoft agreement. The LLMs that brokers and agents call are billed by their own providers.
How to start
Take stock
List the agents, MCP servers and APIs already running, and let the scanners fill the registry.
Set policies
Authentication, access and security rules for agent and LLM traffic, and budgets for model spend.
Expose tools
Turn the APIs agents need into MCP servers with MCP Bridge.
Build one network
One agent network for one process, with fixed paths where the process must not vary.
Watch and widen
Agent Visualizer and alerts in production, then the next process.
What to check before you start
- Data and law: which MuleSoft control plane and region your organization uses, against your data rules and the law that applies to you, such as the UAE's personal data protection law or the GDPR.
- Release level: some capabilities are newer than others, so check that each one you plan to use is generally available.
- Model costs: model providers bill separately, so set budgets before launch.
- Ownership: agree who owns the registry, policies and networks across teams.
How Birbal AI helps
Birbal AI sets up MuleSoft Agent Fabric with engineers based in the UAE, alongside our API integration work. Birbal AI is an independent company, not part of MuleSoft or Salesforce.
Questions and answers
What is MuleSoft Agent Fabric?
MuleSoft's control plane for AI agents, MCP servers and APIs: Agent Registry, MCP Bridge and A2A connectors, Omni Gateway policies with a model proxy for LLM costs, Agent Broker networks and Agent Visualizer.
Is Agent Fabric only for Salesforce agents?
No. MuleSoft says it governs agents regardless of where they are built, and its scanners find agents on Amazon, Google, Microsoft, Databricks, ServiceNow and Kong platforms, among others.
What are Agent Broker and Agent Script?
Agent Broker routes work across agents and tools in an agent network. Since July 2026 brokers are defined in Agent Script, a graph-based language with deterministic and LLM-driven nodes.
How much does MuleSoft Agent Fabric cost?
MuleSoft does not publish a price and asks customers to contact their account team. The LLMs your agents use are billed separately by their providers.
Where do agent networks run?
On CloudHub 2.0, on Anypoint Runtime Fabric or, since 16 November 2026, on Hyperforce for agent networks version 2.0.
Does Birbal AI set up Agent Fabric for us?
Yes, with engineers based in the UAE. Ask for a free estimate and get a scope, timeline and cost range within 48 hours.
Sources
- MuleSoft Agent Fabric (MuleSoft)
- Agent Fabric overview (MuleSoft documentation)
- Agent Fabric release notes (MuleSoft documentation)
- Building agent networks for Agent Fabric (MuleSoft documentation)
- Salesforce launches MuleSoft Agent Fabric, 25 September 2025 (Salesforce)
Birbal AI is an independent company based in Abu Dhabi. Facts about third parties come from the sources above; check them for the latest position.