What Is an Agentic Mesh?
An agentic mesh is an architecture for operating many AI agents, built by different teams or vendors, as one governed network. Agents register their capabilities, discover one another, delegate tasks and share context through common services for identity, access control, observability and policy, rather than through point-to-point integrations. The term names an organizing pattern for an agent ecosystem, not a single product or protocol.
Updated
What is an agentic mesh?
Two sources shaped the term. Eric Broda wrote about "Agentic Mesh" in Towards Data Science in November 2024, defining it as "an interconnected ecosystem that makes it easy for autonomous agents to find each other, collaborate, interact, and transact." McKinsey's report Seizing the agentic AI advantage (June 13, 2025), billed as "A CEO playbook to solve the gen AI paradox and unlock scalable impact with AI agents," set out the variant "agentic AI mesh": "a composable, distributed, and vendor-agnostic architectural paradigm that enables multiple agents to reason, collaborate, and act autonomously across a wide array of systems, tools, and language models—securely, at scale, and built to evolve with the technology." Its authors are Alexander Sukharevsky, Dave Kerr, Klemens Hjartar, Lari Hämäläinen, Stéphane Bout and Vito Di Leo, with Guillaume Dagorret, "representing views from QuantumBlack, AI by McKinsey and McKinsey Technology."
The problem McKinsey describes is scale. Agents, it argues, bring "a new class of systemic risks": "uncontrolled autonomy, fragmented system access, lack of observability and traceability, expanding surface of attack, and agent sprawl and duplication." Components "such as memory stores or orchestration engines" are, it says, "necessary," but "not sufficient" when bolted "on top of existing gen AI stacks." The name follows two earlier ecosystems: Eric and Davis Broda describe the service mesh for APIs and the data mesh for data products as its predecessors.
How an agentic mesh works
McKinsey's model rests on five design principles:
- Composability. "Any agent, tool, or LLM can be plugged into the mesh without system rework."
- Distributed intelligence. "Tasks can be decomposed and resolved by networks of cooperating agents."
- Layered decoupling. "Logic, memory, orchestration, and interface functions are decoupled to maximize modularity."
- Vendor neutrality. Open standards "such as the Model Context Protocol (MCP) and Agent2Agent (A2A) are preferred to proprietary protocols."
- Governed autonomy. Behavior is "proactively controlled via embedded policies, permissions, and escalation mechanisms."
It then names seven capabilities:
| Capability | What it does in McKinsey's description |
|---|---|
| Agent and workflow discovery | A catalog of agents and workflows, for reuse and for policy on their use |
| AI asset registry | Governs system prompts, agent instructions, model configurations and tool definitions, with version control and access policy |
| Observability | End-to-end tracing across agentic and procedural systems |
| Authentication and authorization | Fine-grained access control between agents, systems and models |
| Evaluations | Testing agent pipelines for accuracy and compliance over time |
| Feedback management | Feedback loops that evolve agent configurations |
| Compliance and risk management | Policy controls, compliance agents and guardrails |
In a working system, A2A carries delegation between agents and MCP carries agent access to tools. The mesh adds the organization-wide pieces around them: a catalog of every agent, one identity and permission model, and one trace that follows a request across both.
Why an agentic mesh matters in real-time systems
McKinsey says the mesh supports "multiagent collaboration by allowing agents to share context and delegate tasks." The seven capabilities decide who may act, what gets recorded and how agents are judged. None of them, as listed, says what happens when two agents act on the same state at the same moment. That falls to whatever holds the state, and a mesh design should answer three questions about it:
- Who serializes unmergeable writes? Two agents reserving the last unit of inventory, or drawing on one budget, both pass authorization. Only a transaction or conditional write where the state lives stops both from committing.
- Do agents that share context read the same version of it? Delegation passes references and summaries between agents. If each agent then reads the underlying entity from a different system, they reason about different moments.
- How fast do changes reach running agents? The asset registry versions instructions and tool definitions. When a version changes or a permission is revoked, agents already in flight keep acting on the old one until they reload, and that window has a cost.
These are the failure modes covered in agentic AI architecture: patterns and what breaks in production.
FAQ
Related terms
An agent handoff transfers an in-progress task from one AI agent to another. What has to move with it: the conversation, the task's state and authority to act.
Tool calling, or function calling, lets a language model ask its application to run an external function and return the result as input to its next step.
A fleet control loop observes running AI agents, derives fleet state, adjusts policy, budget or routing, and has agents act on the change while still running.
Multi-writer contention is two or more actors reading and writing the same mutable state in the same short window. It is not multi-tenancy or partitioning.
Unmergeable state is shared mutable state whose concurrent updates can't be combined afterward without breaking a rule, such as a budget, quota or balance.
The blackboard pattern coordinates independent specialists through one shared workspace they read and write, instead of through messages between them.