Mark Kretschmann(@mark_k):Turns out that my "shower thought" about using AI models for harnesses was a bit prescient: This was just turned into reality with the Prime Agent harness by Prime Intellect!

2026.08.06 09:18

Turns out that my "shower thought" about using AI models for harnesses was a bit prescient: This was just turned into reality with the Prime Agent harness by Prime Intellect!

Mark Kretschmann@mark_k

2026.07.21 12:03

Shower thought: the harness around an LLM probably shouldn’t be rigid code. It should itself be an AI model. Today, most agent systems have an intelligent model trapped inside a fairly dumb shell: fixed routing, fixed workflows, fixed retry logic, fixed context selection. A better architecture would use another model as the executive layer. It could decide how to approach the task, which tools or models to use, what context matters, when to verify, when to retry, and when to change strategy entirely. Some boundaries should still remain hard-coded, especially permissions, safety, and validation. But the actual orchestration should be adaptive.