Define the task before choosing a model

I am building my own agent platform, Harness. Model selection is one part of it. First I need a bounded task: which files belong to it, what result do I expect and how can I check it? I set the architecture and rules. Agents write much of the code under my direction.

Check rather than guess

I want an assistant I can rely on as far as possible. That means acknowledging uncertainty and checking assumptions against sources. Another agent can help spot something that was missed. I still need to check the result myself.

Handoffs belong in the workflow

A full context window does not finish the task. context-parachute (CLI / hooks) records the state for the next agent. lore (local, Python) archives the history. Both tools grew out of this need and are available on GitHub under the MIT licence.

A review has limits

Before a merge to main, two models review the commit. A third resolves disagreements. A review can be wrong too. Tests and responsibility for the change still matter.