Defining what an agent is has become easier over the past year as the term has been applied more consistently. An agent is a system that uses a language model to take a sequence of actions toward a goal, where the actions and their sequence are determined by the model based on the current state of the task. This distinguishes agents from single-shot LLM applications, where the prompt goes in and the response comes out, with no loop.
Defining what it means to run an agent well is significantly harder. The infrastructure question — what is the runtime environment for an agent? — is one that the field has not yet answered well, and the gap between the best current answer and the right answer is large.
What a Runtime Needs to Do
A runtime for agents needs to handle several concerns that are entirely absent from single-shot LLM applications. Execution state: the agent's current position in a task, the history of actions taken, the intermediate results accumulated. Scheduling: when does the agent run? Is it event-driven? Periodic? Does it block on a user response? Tool execution: dispatching tool calls to external systems, handling the response, and incorporating it into the agent's context. Failure handling: what happens when a tool call fails, when the model produces an unparseable output, when the execution exceeds its time or cost budget?
None of these concerns exist in isolation. They interact in ways that make them difficult to handle in ad-hoc application code. An agent that loses its execution state when a tool call times out is not a reliable agent. An agent that does not have a coherent strategy for handling model failures will behave unpredictably in production.
The Current State
The current state of agent runtime infrastructure is basically: each team builds their own. LangChain provides some scaffolding, AutoGPT demonstrated the pattern, but the production-grade runtime infrastructure that would let teams deploy agents reliably and manage them operationally does not really exist yet as a first-class product. What does exist is either research code or application-specific implementations with no general-purpose design. This gap is what Superagent, which we backed earlier this year, is specifically working to close — building the runtime infrastructure that makes deployment of conversational agents tractable for production applications.
We expect this to be one of the more important infrastructure categories to crystallize over the next two years. The teams that get the agent runtime abstraction right will be in a position analogous to where application servers were in the late 1990s: foundational infrastructure that every production deployment depends on.