Network-AI adds coordination layer to fill MCP's multi-agent gaps
Jovan Marinovic argues that MCP solves agent-to-tool connections but leaves agent-to-agent state coordination broken, and introduces Network-AI, an open-source coordination layer that adds atomic state updates, conflict resolution, and token budget controls on top of existing frameworks.
Score breakdown
Teams running multiple MCP-powered agents in production should audit their shared state writes — silent overwrites require an explicit coordination layer like Network-AI rather than relying on framework defaults.
- 01Concurrent agent writes to shared state can silently overwrite each other with no error thrown — identified as the #1 failure mode in multi-agent production systems.
- 02Network-AI is an open-source (MIT license) coordination layer built by Jovan Marinovic to sit between agent frameworks and shared state.
- 03All state mutations in Network-AI go through a `propose → validate → commit` cycle to ensure atomic, conflict-free updates.
Jovan Marinovic's post on Dev.to identifies a production failure mode that most multi-agent framework discussions overlook: shared state corruption from concurrent agent writes. While frameworks like LangChain, AutoGen, and CrewAI each handle individual agent capabilities well — chains, conversations, and roles respectively — none of them coordinate shared state when multiple agents operate simultaneously. Marinovic illustrates the problem with a concrete timeline: two agents reading the same context version at `0ms` and `5ms`, then both writing back, with the second write silently overwriting the first at `15ms` with no error raised.
To solve this, Marinovic built Network-AI, an open-source MIT-licensed coordination layer that sits between agent frameworks and shared state.
To solve this, Marinovic built Network-AI, an open-source MIT-licensed coordination layer that sits between agent frameworks and shared state. Rather than allowing direct writes, every mutation goes through a `propose → validate → commit` cycle via `networkAI.propose()`, which validates against concurrent proposals, resolves conflicts automatically, and commits atomically. The project claims support for 14 frameworks including LangChain, AutoGen, CrewAI, MCP, A2A, and OpenAI Swarm.
Beyond atomic state management, Network-AI adds token budget controls per agent to prevent runaway costs, role-based permission gating across agents, and a full audit trail of each agent's actions. Marinovic frames Network-AI as complementary to MCP rather than a replacement — MCP handles the agent-to-tool layer, while Network-AI handles the agent-to-agent coordination layer — positioning the two together as a complete production stack for multi-agent systems. The project is available on GitHub and has an associated Discord community.
Key facts
- 01Concurrent agent writes to shared state can silently overwrite each other with no error thrown — identified as the #1 failure mode in multi-agent production systems.
- 02Network-AI is an open-source (MIT license) coordination layer built by Jovan Marinovic to sit between agent frameworks and shared state.
- 03All state mutations in Network-AI go through a `propose → validate → commit` cycle to ensure atomic, conflict-free updates.
- 04Network-AI claims support for 14 frameworks, including LangChain, AutoGen, CrewAI, MCP, A2A, and OpenAI Swarm.
- 05Features include token budget controls per agent, role-based permission gating, and a full audit trail.
- 06