Agent Mesh: How to Build an Enterprise Platform for AI Agent Collaboration

Watch the full session · 25 min

At COLLIDE 2025, Grid Dynamics Search and AI Architect Stas Stolpovskiy shared how the company built an internal agent mesh. He explained what a central team must control, why platforms need a few strict rules and lots of freedom, and why he thinks MCP falls short for agent-to-agent work.

Getting one AI agent to work on a laptop is easy. Getting hundreds of agents built by different teams to work together across an enterprise is a platform problem, Stas Stolpovskiy, Search and AI Architect at Grid Dynamics, told attendees at COLLIDE Data + AI Conference 2025 in Atlanta. In a session sponsored by Grid Dynamics, he described the architecture behind the company's internal agent mesh and the lessons from two years of building agentic systems.

What is an agent mesh?

Stolpovskiy explained the idea with a freelancer marketplace. A few years ago, a company that needed a website widget built might post the job on a freelancing site, pick a contractor, exchange files and messages, and track progress, all inside one platform that also guaranteed privacy, contracts and transparency.

"After six years this is exactly how you're building the agent mesh," he said. Agents now take on the simple tasks freelancers once did, and the enterprise needs the same components for discovery, communication, safety, control and transparency.

He distinguished three levels. A simple agent is a prompt with a few tools that does one job. Agentic AI is multiple agents with their own autonomy that know about each other and work together. An agent mesh adds the platform that connects agents from many parties, each using different tools and connecting in different ways.

Why do enterprises need an agent platform?

Stolpovskiy gave four reasons enterprises can't rely on a simple off-the-shelf tool:

• Breaking silos: connecting many agents and data types securely brings data to employees and customers more widely.
• Scale: a system for 5,000, 10,000 or 50,000 employees needs different infrastructure than one on a laptop.
• Transparency: audits and tracing must work for everyone, including public companies looking for vulnerabilities.
• Safe innovation: the most important reason, he said. Just two weeks earlier, different models had topped the leaderboard. "The market is changing insanely fast," he said, so platforms must let teams plug in and test new components securely.

Should you build an AI platform top-down or bottom-up?

Grid Dynamics has two large customers taking opposite approaches, Stolpovskiy said. In the top-down model, leadership defines the platform and builds guardrails, retrieval systems and connectors before there is a use case. The common result is a well-built platform that nobody uses.

In the bottom-up model, a team ships a prototype for a use case many people need, shows a return and then tries to scale it. Each has trade-offs, and Stolpovskiy said his instinct is that the middle path works best: don't build a platform without users, and don't build a use case without thinking about how it will scale.

Either way, he said, teams face the same architecture questions: availability and durability, evaluation, agent experience, governance and security, integrations with data sources and APIs, the runtime where agents run, and agent state over a lifecycle that can now last for days.

What should a central AI team control, and what should it leave open?

Stolpovskiy split the platform into two parts. Observability, a unified user experience, security and integrations with internal data must be controlled centrally. Everything else changes so fast that the platform team should guide rather than mandate, offering reference implementations.

He warned against creating yet another standard and suggested a different model. "Be USB-C. Everyone loves USB-C," he said. The rules should be few and strict: agents are registered and discoverable, share a flexible common contract, support end-to-end tracing and handle security token delegation. Beyond that, teams can use LangGraph, ADK or whatever framework they prefer.

What works, he said, is a simple, extendable and backward-compatible protocol, so an agent built a year ago keeps working after its creators have moved on. The pitfalls: letting teams adopt unproven frameworks straight from GitHub, failing to plan for scale, and spending six months building a platform before getting user feedback.

Is MCP the right protocol for an agent mesh?

Stolpovskiy made what he called a claim many would argue with: inside an agent mesh, MCP restricts teams compared with agent-level protocols. "MCP was designed to be the end interface for the agent," he said. It was never meant to solve many use cases under the hood.

He pointed to the Agent2Agent (A2A) protocol as a better fit for connecting agents because it was built for that purpose, with room for security, capabilities and interfaces. He showed agents that return forms or diagrams rather than text. "The agent is deciding what kind of experience you will get," he said, and the protocol must allow UI widgets to be reused as tools without breaking older agents.

How do agents resolve conflicting data?

An audience member asked what happens when two agents return contradictory data from different sources. In a mesh, Stolpovskiy said, the orchestrator works with agent outputs, not raw data. A good agentic system includes a step that critiques results, and it can ask an agent why its answer differs, using evidence such as timestamps to resolve the conflict.

He closed with a four-step cycle: build a solid platform, promote it so users start building on it, roll it out quickly to get feedback, then analyze and improve. Grid Dynamics runs the system internally first. "We're saying drink our own champagne," he said.

Hear more agentic AI architects at COLLIDE 2027

COLLIDE, the largest AI conference in the Southern U.S., brings together more than 1,000 enterprise data and AI leaders. COLLIDE 2027 takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center in Atlanta. Registration is coming soon.

Watch: Inside the Agent Mesh: Building an Enterprise Scale Platform for AI Collaboration Stas St on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights