What's this about?
A framework for evolving environment-specific agent harnesses while keeping model weights fixed. What counts as part of the harness here is unusually well enumerated: prompt and task framing, tool interfaces, skills, MCP-backed providers, subagent structure, and agent-loop configuration. 20 upvotes.
Why does "weights fixed" matter?
Because it changes who can do this. Training a model requires a lab; evolving a harness requires a laptop and patience. If the harness dominates capability — which three separate papers this month have argued — then adapting the harness to your environment is the highest-leverage thing available to anyone who isn't a frontier lab. "Enterprise environments" in the title is the point: your company's tools, conventions and constraints are not in anyone's training data, and the harness is where that knowledge goes.
Why does this matter to you?
- Third harness-evolution paper in two weeks, after JIT-Agent and behaviour-aware verification — this is now a research front.
- The enumeration of what a harness contains is genuinely useful as a checklist for auditing your own setup.
- It is the practical answer to "we can't train models" — you can still evolve everything around them.