Single-step agents were the proof of concept. A user provides a task, an LLM decides what tool to call, the tool executes, and the result comes back. That loop demonstrated something important: that language models can reason about actions, not just text. But it also exposed a ceiling. Single-step agents are brittle, shallow, and poor at tasks that require any meaningful planning horizon.
What comes next is the compound agent — an architecture where the primary agent does not execute tasks directly but instead orchestrates a network of specialized subagents, each responsible for a specific capability or domain. The compound agent reasons about what needs to happen, decides which subagent to engage, monitors their execution, and synthesizes the results.
The Delegation Problem
Building compound agents surfaces a problem that single-step architectures never encounter: trust and delegation. When one agent spawns another, several questions immediately arise. Does the spawned agent have the right permissions for the context it is being given? What happens when a subagent fails partway through a multi-step task? How does the orchestrator know when to intervene versus when to let a subagent retry?
These are not LLM problems. They are distributed systems problems. And the infrastructure to handle them — agent runtimes with robust state management, execution tracing, and failure recovery — is still early. The companies building this substrate are some of the most interesting in the portfolio right now.
Composition as the Real Moat
The deepest moats in the agentic stack will not belong to the companies that build the best individual agents. They will belong to the companies that make agent composition tractable. Composition means something specific: the ability to wire together agents with different capabilities, different context, and different execution semantics, without requiring the developer to hand-code every interaction.
This is analogous to what happened in the microservices era. The real value was not in any individual microservice — it was in the tooling that made orchestrating dozens of microservices manageable. Kubernetes, service meshes, distributed tracing. We expect to see an analogous stack emerge for agentic systems over the next 18 months. We intend to be early shareholders in several of those companies.
The window to build this infrastructure is not wide. Application teams are already running compound agent architectures in production, duct-taped together with custom code. The moment the right abstractions appear, adoption will be fast.