liveBook · Manning

Agent mesh

As explained in Architecting for Autonomy

A coordination architecture in which multiple agents interact through supervisory, brokered, and peer-to-peer structures under defined contracts, authority, paths, and evidence obligations.

In a quality investigation, several specialist agents may gather evidence, assess supplier history, evaluate regulatory concerns, and monitor changing signals. Their contributions must lead to a defensible containment recommendation without letting collaboration silently expand anyone’s authority. An agent mesh makes those collaboration boundaries explicit. section 6.5 section 6.5.2

An agent mesh is a governed structure for collaboration among specialist and supervisory agents. It defines permitted paths, interaction contracts, delegated authority, escalation routes, and evidence obligations. It is one architectural realization of a governed multi-agent system, not unrestricted agent-to-agent communication. section 6.5 section 6.5.1

Communication protocols can carry messages, but they do not decide who may send them. Runtimes, endpoints, and registries carry collaboration; permitted-path definitions, policy services, durable workflow or case state, and evidence records realize the governance outside the runtime. Governed wrappers, events, and human re-entry paths connect agents to systems and people without themselves becoming agent participants. section 6.5.1 section 6.9

Table 6.6 Agent Mesh pattern card
Attribute
Definition & Requirements
Mandate
Design collaboration between agents as an explicit, governed structure, not as an emergent chain of prompts, tool calls, or informal handoffs.
Scope
Applies where multiple specialist or supervisory agents must coordinate to achieve an outcome across systems, domains, or decision surfaces, including cases where their collaboration uses brokers, governed tools, events, or human handoff points.
Operating envelope
Valid when each participating agent has a defined role, permitted interaction paths, coordination rules, escalation routes, and evidence obligations before runtime. The mesh must make clear which agents may initiate, delegate, approve, execute, monitor, reconcile, or escalate work.
Boundary of authority
No agent may expand its authority by delegating to, prompting, or composing with another agent. Authority must remain traceable to the originating contract, decision surface, or delegated capability. Coordination does not create new authority; it only routes authority that has already been granted.
Failure mode
If agents conflict, loop, duplicate work, lose context, exceed authority, or produce incompatible outcomes, the mesh must fail through containment, arbitration, escalation, or human handoff rather than allowing unresolved coordination to continue.
Evidence & auditability
The mesh must record the originating intent, participating agents, delegated tasks, authority basis, messages or events exchanged, decisions made, conflicts resolved, handoffs triggered, and final outcome. Reviewers must be able to reconstruct not only what each agent did but also how the collaborative outcome was produced.

The pattern card sets out the mesh’s mandate, operating envelope, and authority boundaries. section 6.5.1

Three coordination structures can coexist within one mesh:

  • Supervisor: one agent holds the outcome contract, delegates bounded tasks, and reconciles results. It provides a single explainable point, but sits on every path and may accumulate business logic. section 6.5.2
  • Brokered: an intermediary mediates access to shared resources or enforces policy without contributing domain reasoning. A broker that starts deciding whose request matters more is drifting into orchestration. section 6.5.2
  • Peer-to-peer: specialists exchange a bounded signal directly when routing through a supervisor or broker would be too slow or too broad. section 6.5.2

Quality-investigation mesh

A supervisor delegates evidence gathering, supplier history, regulatory assessment, and signal reclassification, then reconciles the results into a containment recommendation. A production data broker controls specialist access to manufacturing data through governed wrappers. If detection reclassifies a suspected defect above a defined confidence threshold, it can signal evidence assembly directly so work does not continue under the old assumption. Decisions requiring retained human authority return through the relevant human holders. section 6.5.2 section 6.5.1

Test the coordinating layer whenever it takes on a new responsibility: if replacing it would change business policy, authority, evidence obligations, or system permissions, too much architecture has leaked into coordination. A healthy replacement should change only sequencing, routing, and handoff mechanics. No agent should gain authority merely by delegating to or composing with another agent. section 6.5.1

How Architecting for Autonomy covers it

How deep it goes
Defines governance requirements, compares three coordination structures, and provides a test for coordination drift.
Examples
Uses a quality investigation with a supervisor, specialist agents, a production data broker, and a direct defect-reclassification signal.
What you can do after
Evaluate a proposed mesh’s collaboration paths, authority boundaries, and coordination layer for governance drift.
How it connects
Connects multi-agent systems with conversational contracts, governed tool wrappers, event-mediated coordination, orchestration, and human re-entry.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Architecting for Autonomy. The section links point to the parts of the book it draws on.