Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-16 · high
That means teams running self-improving harnesses can generate proposals on a cheap model and reserve the expensive one for judging and applying the change.
Why now: Weng posted the survey July 4, and it's the harness-economics number surfacing in mid-July coverage of self-improving agents.
Harness-updating capability doesn't improve from Qwen2-32B up through Opus 4.6, per a survey by Lilian Weng posted July 4. Weng was testing whether a pricier model writes better proposals for updating a coding harness. A harness is the rules and scripts that let an agent revise its own workflow. The result: no meaningful gain. A 32-billion-parameter model proposes changes about as well as the priciest model in the test.
That's a real number for anyone running a harness with a self-improvement loop. Burning frontier-model tokens to generate those proposals buys nothing extra. Weng's survey puts the bottleneck somewhere else: utilization, the judging and applying of a proposed change, not the writing of it.
The split that follows: generate cheap, judge and apply expensive. Run proposal generation on something in the Qwen2-32B class. Save the frontier-model budget for the step that decides whether a change is safe to merge, then executes it. Paying Opus-level rates for the generation step buys a skill that plateaus early.
I run a harness that proposes its own fixes, and this is the split I'd bet on. The expensive model should sit at the judge step, not the writer's desk.
Each link below shares sources, entities, or timing with this story.
Claude Code uses Opus / Shared entity: Opus / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (frontier, model, opus).
Claude Code uses Opus / Shared entity: Opus / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (capability, frontier, model, opus).
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (expensive, model, opus).
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (doesn, model).
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (model, opus).
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (model, opus).
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (frontier, model).
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (model, opus).