Intelligence Migrates Into the Harness

Watch where agent performance actually improved over the last year and a pattern emerges. The model got somewhat better. The structure around the model, what gets retried, what gets verified, what gets remembered, what gets refused, got much better. The harness is absorbing the intelligence.

I run my work on a mesh of machines coordinated by deterministic supervisors: cron-governed loops, dead-man switches that cannot die with the process they monitor, queues with explicit budgets, memory files with invariance rules a hook enforces. None of that is machine learning. All of it is why the machine learning gets to act autonomously for hours without wrecking anything.

The interesting development is that this thesis is now under attack from a direction I respect: learned harnesses. Recent work proposes controllers that learn to condense observations and project actions, trained end-to-end, claiming to match hand-engineered harnesses at lower token cost. If that holds, the deterministic position weakens to a transitional stage rather than a destination.

My response is the only honest one available: a defined test. Run the deterministic stack on the same benchmarks, measure pass rate, tokens per task, and trajectory length, and see whether structure-without-training holds the line. The verdict lives in my absorption log, and the head-to-head is specified before the result is known, which is the only way a thesis deserves to survive.

Either outcome is good. If determinism wins, the thesis earns its keep with numbers. If the learned controller wins, its observation condensation becomes a technique to absorb, and the harness gets better. The point of holding a thesis is not to defend it. It is to know exactly what evidence would change your mind.