Build the Infrastructure, Not the Harness

If you asked me whether we could build a better model or a better harness than a frontier lab, the honest answer is no. Any startup that claims otherwise is either lying or about to burn a lot of capital proving the point. But ask a different question, whether we can build better infrastructure for running harnesses at enterprise scale, and the answer flips. That is not what the labs do. It is not what they are staffed for, and it is not where their attention goes.

The distinction matters because the two layers reward completely different work. The harness layer moves fast and commoditizes fast. Prompts get copied and agent loops get replicated within weeks. Worse, any magic you inject into an agent loop you do not own becomes debt: the agents are now good enough that they do not need determinism grafted on, and every proprietary intervention is something you eventually have to unwind. The infrastructure underneath is brutally hard engineering: sandboxing, environment provisioning, isolating secrets so an agent can use credentials without ever reading them, running a harness you do not own inside the customer's own cloud, air-gapped if the customer demands it. The labs ship SaaS, not self-hosted runners. That difficulty is not a bug in the plan. It is the plan.

A second force pushes the same direction, strategic rather than technical. No serious CTO wants to be locked into one vendor's models or one vendor's harness. The leaderboard changes every quarter, and teams act on it: companies that standardized their entire engineering org on one coding harness have already moved everyone to another, sometimes inside a single renewal cycle. Developers push the same way from below: they do not want a wrapper around the coding agent, they want raw access to the standard harnesses, unmodified and side by side, so the work invested in one carries over to the next. The layer that wins is the one that stays agnostic, the one where swapping the harness underneath is a configuration change instead of a migration.

This is the same lesson the harness layer's missing moat has been teaching for a year. The labs own the models. The harnesses converge. What is left, and what stays hard, is the infrastructure that makes any of it deployable inside a real enterprise.

So when you evaluate an agent vendor, or your own build plan, ask which layer they actually own. If the answer is a prompt and a wrapper, the labs will eat it. If the answer is infrastructure the labs will never bother to build, that is a position worth holding.