MuleSoft Agent Fabric: What It Is and How to Start
18 November 2026 · Leadex Systems
Most large companies now have AI agents in more places than anyone can list: Agentforce in Salesforce, a copilot in the office suite, agents built on Amazon Bedrock or Google, and a few custom ones calling internal APIs. Each team chose its own tools, and nobody has the full picture of which agent can do what, on which systems, and at what cost.
MuleSoft Agent Fabric is MuleSoft's answer to this agent sprawl. This article explains what it is, what its parts do, how MCP and A2A fit in, what it will not do for you, and a 90-day plan to start. It is based on the public MuleSoft documentation (November 2026) and on our webinar with banking teams on the same topic.
What is MuleSoft Agent Fabric?
Agent Fabric is a set of Anypoint Platform capabilities to discover, orchestrate, govern and observe AI agents and the tools they use, whichever vendor built them. It builds on products many MuleSoft customers already run: Anypoint Exchange as the catalog, the MuleSoft gateway and API Manager policies for enforcement, Anypoint Monitoring for metrics, and CloudHub 2.0 or Runtime Fabric to run agent brokers. See the Agent Fabric overview in the MuleSoft docs.
MuleSoft says Agent Fabric is generally available, with new capabilities rolling out every month. Pricing is not published; it is quoted by your Salesforce account team (MuleSoft product page).
The four parts of Agent Fabric
1. Discover: Agent Registry
One inventory of your agents, MCP servers, APIs and related services, including curated public MCP servers. Scanners watch supported platforms (MuleSoft names Amazon, Google, Microsoft, Claude, Databricks and Kong, among others) and register what they find, so the inventory does not depend on teams remembering to fill in a spreadsheet.
2. Orchestrate: Agent Broker
A broker receives a task and routes it to the agent, or to another broker, best suited to handle it. Brokers run on CloudHub 2.0 or Runtime Fabric. Salesforce's architects recommend a hierarchy of brokers by business domain rather than one flat orchestrator, and a short allowed list of tools per broker: in their experience most LLMs handle about 20 to 25 tools before accuracy drops.
3. Govern: policies, security and cost
Agent and tool traffic goes through the MuleSoft gateway (Omni Gateway in the current docs, Flex Gateway in earlier material), where policies apply as they do for APIs today. On top of that the docs list governance strategies, rogue agent detection, an agent kill switch, and cost controls: a model proxy with routing and caching, and budgets per team or agent.
4. Observe: Agent Visualizer
A map of which agents call which MCP servers and models, with latency, throughput and error rates. It answers the questions a risk or audit team asks first: what is connected to what, and who did what.
MCP and A2A: how your APIs become agent tools
Agent Fabric uses two open protocols. MCP (Model Context Protocol) is how an agent calls a tool, such as "get account balance". A2A (Agent2Agent) is how agents hand tasks to each other. MuleSoft offers four building blocks:
- MCP Bridge turns an existing API or SaaS system into an MCP server without custom code.
- MCP Connector runs an MCP server or client inside a Mule application, when you need logic, orchestration or data mapping behind the tool.
- A2A Bridge makes an agent that does not speak A2A, including many Agentforce agents, part of the network without rebuilding it.
- A2A Connector lets a Mule application act as an A2A server, client or both.
The practical point: if your core systems already sit behind well-designed, secured MuleSoft APIs, most of the work of giving agents safe access is already done. Agents should reuse those APIs instead of getting their own direct connections.
Does Agent Fabric work with non-Salesforce agents?
Yes, by design. Salesforce's architect deep dive describes Exchange as holding A2A agents, MCP servers and LLM providers "from any source, including third-party, Agentforce, and custom MuleSoft agents". One condition: the broker orchestrates only A2A-compliant agents, so agents that do not support A2A go through the A2A Bridge first.
What Agent Fabric will not do for you
A platform gives you the controls. It does not make the decisions. From our integration work, these gaps matter most:
- It won't fix weak APIs or data. An agent calling a poorly designed API with inconsistent customer data gives poor answers, just faster.
- It won't assign owners. Every agent needs a named business owner who decides what it may do and answers for it.
- It needs infrastructure. Governed traffic runs through gateways (the architect guide uses one for incoming and one for outgoing traffic), and brokers need a CloudHub 2.0 space or Runtime Fabric.
- Observability is still growing. The architect guide lists OpenTelemetry tracing and per-agent token cost tracking as future capabilities, so plan how you will cover them for now.
- Humans stay in the loop. Quarantining a suspicious agent always needs human review, and budgets are approximate: usage can briefly exceed a limit.
Five integration lessons that apply to agents
Banks learned these the hard way with ESBs and point-to-point integrations. Agents repeat the same mistakes faster:
- Keep an inventory. You cannot govern what you cannot list.
- Access only through APIs. No direct database or screen-scraping access for agents.
- Security from day one. Identity, least privilege and audit logs before the first pilot, not after.
- An owner for every agent. Orphaned integrations were expensive. Orphaned agents are worse, because they act.
- Clean data first. One customer view beats ten agents guessing from ten systems.
A 90-day plan to start with Agent Fabric
Days 1–30: get the picture (no budget needed)
- List the agents, copilots and MCP servers already in use, with an owner for each.
- Pick one business scenario with clear value and limited risk, such as a service request that today touches three systems.
- List the APIs that scenario needs and check which exist, and how well they are secured.
Days 31–60: connect and govern
- Register the agents and tools for the scenario in Exchange and the Agent Registry.
- Expose two or three existing APIs as MCP tools with MCP Bridge, and close any API gaps with the MCP Connector.
- Put gateway policies in place: authentication, rate limits, and logging that your security team signs off.
Days 61–90: pilot and decide
- Run the scenario through an Agent Broker with real users and a human approval step.
- Watch the Agent Visualizer and the gateway metrics: errors, latency, cost and what the agents actually called.
- Decide with evidence whether to scale, change the scenario or stop.
How Leadex Systems can help
Agent Fabric is new, and so is everyone's experience with it. What we bring is the integration and API governance work it builds on, mostly in financial services. We usually start with a MuleSoft assessment of the APIs your agents would use, then a scoped pilot on one scenario. See AI agent integration with Agentforce and Agent Fabric, or start small with our MuleSoft consulting services package.
Russian-speaking teams can also watch our webinar on AI agents in banks and MuleSoft Agent Fabric (in Russian).
Frequently asked questions
What is MuleSoft Agent Fabric?
A set of MuleSoft Anypoint Platform capabilities to discover, orchestrate, govern and observe AI agents and the tools they use, across vendors. Its main parts are the Agent Registry, Agent Broker, gateway-based governance and the Agent Visualizer.
Is MuleSoft Agent Fabric generally available?
MuleSoft says Agent Fabric is generally available, with new capabilities released every month. Check the MuleSoft documentation for the status of individual features.
How much does MuleSoft Agent Fabric cost?
MuleSoft does not publish Agent Fabric pricing. It is quoted by your Salesforce or MuleSoft account team. Budget also for the gateways and the CloudHub 2.0 or Runtime Fabric capacity that brokers run on.
Does Agent Fabric only work with Agentforce?
No. It is designed to manage agents from any source, including third-party platforms and custom agents. The Agent Broker orchestrates A2A-compliant agents; others join through the A2A Bridge.
What is the difference between MCP and A2A in Agent Fabric?
MCP (Model Context Protocol) is how an agent calls a tool, such as an API. A2A (Agent2Agent) is how agents pass tasks to each other. Agent Fabric supports both: MCP Bridge and MCP Connector for tools, A2A Bridge and A2A Connector for agents.
Do we need new APIs for AI agents?
Usually not. Agents should reuse the governed APIs you already run; MCP Bridge can expose them as agent tools without custom code. New APIs are only needed where a scenario has gaps.