The field of multi-agent systems has been dominated for the past two years by frameworks — libraries and abstractions that try to make it easier to wire together multiple LLM agents. LangGraph, AutoGen, CrewAI, and a dozen others. The interest is understandable. Frameworks lower the barrier to experimentation. Developers can stand up a two-agent pipeline in a day and see what happens.
But frameworks are the wrong level of abstraction for the real problems in multi-agent systems. They solve the easy problem — how do I define the topology of agents and describe how they communicate — while leaving the hard problems entirely unaddressed.
The Hard Problems
The hard problems in multi-agent systems are state, trust, and failure. State: how does a multi-agent system maintain consistent shared state across agents that may be executing concurrently, potentially on different machines, potentially against different data sources? Trust: when agent A spawns agent B, what permissions does agent B have, and how are those permissions scoped to the specific task at hand rather than inherited globally? Failure: when an agent in the middle of a complex task fails, what is the recovery path? Does the orchestrator retry from scratch? From a checkpoint? With a different agent?
None of these are questions that frameworks answer. They are questions for infrastructure — for the runtime, the state store, the execution engine that sits beneath the framework layer. The companies that build that infrastructure will be more durable than the frameworks built on top of them.
The Historical Pattern
This plays out the same way every time a new distributed systems paradigm emerges. First come the frameworks — Rails for web, Django for web, dozens of microservices frameworks. Then come the real infrastructure companies — the ones that solve the underlying problems the frameworks glossed over. Kubernetes did not compete with microservices frameworks; it became the substrate they ran on.
We expect the same pattern in multi-agent systems. The current frameworks will drive adoption and surface the real pain points. The infrastructure companies that emerge to solve those pain points will be the lasting businesses. We are actively sourcing in this space and expect to make investments here in 2026.