Agent Fabric & Agent Network Governance (Course #agent-fabric)

Programme every line is something you do

5 modules · 15 topics · in course-flow order

1. Module 1

   Inventory the agent estate

   Objective. Diagnose the unregistered agent estate and produce a Portfolio catalog entry for every agent, MCP server and model proxy in scope.

   • Register agents, MCP servers and LLM proxies in the Anypoint Exchange portfolio, each entry carrying owner, environment and tier metadata
   • Run agent scanners against Amazon Bedrock, Google Vertex AI and Microsoft Copilot, reconciling what they discover against a declared inventory of 38 agents
   • Gate the catalog before any orchestration starts: ≥ 95% of production entries name an owner and state a data-residency zone
2. Module 2

   Agent networks as code

   Objective. Construct an agent network from a versioned agent-network.yaml and deploy it to CloudHub 2.0 as a reviewable artefact.

   • Declare the registry and context sections of agent-network.yaml — agents, agent brokers, LLMs, MCP servers, policies — in Anypoint Code Builder
   • Split reasoning from control flow in Agent Script: orchestrator and subagent nodes for judgement, executor, router and echo nodes for everything else (target: ≥ 60% of graph nodes deterministic)
   • Publish assets to Exchange and deploy to a CloudHub 2.0 private space through the Anypoint CLI Agent Fabric plugin (target: reproducible rebuild in ≤ 15 min)
3. Module 3

   Speak A2A, expose MCP

   Objective. Construct the protocol surface of the network — A2A-compliant Mule apps and MCP servers — and measure the context cost of every tool exposed.

   • Make Mule applications A2A-compliant with the A2A Connector and publish one accurate agent card per agent, pinned to A2A 1.0
   • Turn existing APIs into MCP servers with the MCP Bridge, exposing only the tools an agent actually needs (ceiling: 12 tools per server)
   • Cut context spend with MCP Tool Mapping and MCP Tools Progressive Disclosure (target: −30% prompt tokens per task versus a full tool listing)
4. Module 4

   Policy on every agent hop

   Objective. Design the Omni Gateway policy set for agent traffic and prove enforcement on both A2A and MCP hops.

   • Front every A2A agent and MCP server with an Agent and Tool Instance in API Manager (target: 100% of production agents, 0 direct upstream routes)
   • Chain A2A PII Detector, A2A Schema Validation and MCP PII Detector, then prove each with negative tests carrying seeded patient identifiers across 5 PII categories
   • Bound blast radius with MCP Global Access allow/block rules and MCP Attribute-Based Access Control: every write tool blocked by default, at most 3 exceptions per network, each with a named approver
5. Module 5

   Bound the runaway network

   Objective. Evaluate agent-network behaviour against its failure modes and set the limit that stops each one.

   • Trace a failed task from broker to subagent to MCP tool in Agent Visualizer and Anypoint Monitoring (lookback 30 minutes to 30 days, default 24 hours)
   • Cap runaway reasoning in the network spec — max_number_of_loops (default 25), max_consecutive_errors, task_timeout_secs — sized against a 400 ms p95 routing budget
   • Alert on what Agent Visualizer actually exposes: agents outside policy coverage and failed-request rate above 2%, with a runbook entry written for each

References & Official Documentation every claim above has a source

Module 1

• View Your Portfolio OverviewMuleSoft Docs

Module 2

• Agent Script ReferenceMuleSoft Docs
• Define Your Agent NetworkMuleSoft Docs

Module 3

• Create MCP ServersMuleSoft Docs
• Using A2A Protocol in Agent NetworksMuleSoft Docs

Module 4

• Adding an Omni Gateway Agent or Tool InstanceMuleSoft Docs
• Omni Gateway Agent PoliciesMuleSoft Docs

Module 5

• Monitoring Agentic AssetsMuleSoft Docs

More in Architecture