Agents Are Software, and Software Needs a Factory

There is a framing problem in the market. People talk about agent harnesses as if the harness is the interesting part. It is not. The interesting part is the factory.

An agent harness is a runtime wrapper. It takes a prompt, some tools, maybe some context, and executes a task. That is table stakes. Every serious engineering team can build one. Many already have — the recent wave of companies publicly disclosing their internal agent infrastructure proves this. Block had teams building internal harnesses before they laid those teams off. Stripe has been doing similar work. Ramp, DoorDash, and dozens of others are in various stages of the same journey.

The factory is something different. The factory is the execution engine and the sandboxing system. It is the orchestration layer that handles long-running, resumable work. It is the context management that survives across sessions. It is the model translation infrastructure that lets you start a task on one model and continue it on another without losing state.

That last one sounds simple. It is brutally hard. It requires session translation systems that handle the impedance mismatch between different model APIs, different context window sizes, and different tool-calling conventions. The harness is a weekend project. The factory is a multi-year platform investment.

This is where the real defensibility lives. Not in the agent definition. Not in the prompt. In the ugly, complex infrastructure that makes agents actually work in production. I've argued elsewhere that the framework trap catches every company that mistakes the wrapper for the product. The companies building factories will quietly become the layer underneath the next decade of agent applications — and most of the customers will never know their name.