AI Agents Are the New Microservices, and the Custom-Framework Graveyard Is Filling Up Fast

Every engineering team in 2026 is building its own AI agent framework. I have watched at least a dozen of these projects kick off in the last six months alone. By mid-2027, the vast majority will be abandoned, the engineers who built them will be gone, and the companies will be paying down technical debt on a framework nobody outside the original team ever understood.

I have seen this movie. It was called microservices. It ended with a graveyard of custom service meshes and three platforms that actually won. The agent space is about to do the exact same thing.

The Microservices Bill, Recapped

Between 2014 and 2018, every team with more than ten engineers decided they needed their own service mesh. Custom retry logic, circuit breakers, service discovery, observability — all built because the off-the-shelf options were immature. By 2020, the consensus had landed: Istio, Linkerd, and Envoy won. The custom meshes got rewritten or quietly abandoned. The teams that built them spent two to three engineer-years maintaining infrastructure that delivered no product value.

The winners were not the teams that built the most sophisticated custom meshes. They were the teams that waited, picked a platform that consolidated, and shipped product on top.

The Agent Framework Count, June 2026

The list of frameworks competing to be the agent orchestration layer of your stack: LangChain, LangGraph, LlamaIndex, CrewAI, AutoGen, smolagents, OpenHands, Goose, Mastra, Pydantic AI, Letta, Atomic Agents, Vertex AI Agent Builder, Bedrock Agents, Azure AI Foundry Agents, OpenAI Agents SDK, Anthropic's agent SDK, AWS Strands, Google ADK, Haystack, Semantic Kernel, Rivet, Flowise, Langflow. That is twenty-four, all launched in the last 24 months, all claiming the same abstraction layer for AI agents.

There is no universe in which twenty-four survive. There is not even a universe in which ten survive. We are in the maximum-entropy phase, and the consolidation is going to be brutal.

What the Winners Will Look Like

History gives us a clear template. The agent space will consolidate to three winners, probably two. The first slot goes to the platform that ships the most production deployments fastest — my money is on the SDK shipped by whichever frontier lab you trust most, because agents will run on that lab's models first and the path of least resistance wins. The second slot goes to the framework with the strongest type system and the most boring abstractions — Pydantic AI and LangGraph are the current frontrunners. The third slot goes to whoever owns the deployment story for the enterprise long tail — likely a cloud-vendor agent runtime like Bedrock Agents or Vertex Agent Builder, not an open-source project.

The custom framework your team is building right now is the 2016 homegrown service mesh in disguise. Two years of "we own our stack" pride, followed by three years of "please can someone rewrite this on the platform we should have used from the start."

What You Should Actually Do

Stop building an agent framework. You do not need one. Pick one of the two or three serious platforms, build your agent logic on top of it, and treat the framework as infrastructure you consume, not differentiate on. The companies that win the agent era are going to win on product, workflow, and data — not on the framework they hand-rolled.

The teams that will look smart in 2028 are the ones that picked a platform in 2026 and shipped product. The teams that will look embarrassed are the ones that spent eighteen months building a custom framework to learn nothing the open-source projects had not already figured out.

The Take

The agent framework market will consolidate to two or three winners the same way the service mesh market did. Every custom agent framework being built right now is a 2016 homegrown service mesh in disguise. The bill arrives in 2027, paid in engineer-years you will never get back. Pick a platform. Ship product. Let someone else's team own the framework.

— Mr. Technology

Posted June 4, 2026. The microservices-to-agents parallel is not a metaphor, it is a forecast.