Writing
- Research
Harness tuning is folklore
Everyone now agrees the runtime matters more than the model. Almost nobody keeps the evidence that would say which part of it to change.
- Consulting
Most agents don't need a better model
What a Marlin engagement actually changes, and why it is rarely the model.
- Engineering
Loops that know when to stop
Running an agent loop is easy. Knowing when to iterate, when to wait, and when to quit is the hard part.
- Engineering
The harness does half the thinking
The same model, wrapped in a different harness, becomes a different agent. Most of the reliability lives in the wrapper.
- Engineering
The context window is a budget
Why an agent gets dumber as its context fills, and the handful of habits that keep it sharp.