Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-16 · high
Qwen2-32B proposes agent harness fixes about as well as Opus 4.6 does, which means the expensive model is wasted on the wrong step.
Why now: Weng's harness engineering survey, published July 4 and covering work through that date, appeared in the July 16 briefing.
Lilian Weng's new survey on harness engineering makes a specific claim: the ability to propose improvements to an agent's harness, the scaffolding around a model that handles tool calls, retries, and state, is flat across model sizes. Qwen2-32B generates harness-edit proposals about as well as Opus 4.6 does.
That's a cost problem hiding as a research finding. If proposal quality doesn't scale with model size, then routing every harness-edit task to a frontier model spends capacity where it buys nothing. The bottleneck Weng identifies isn't generation, it's utilization, meaning the hard part is judging which proposed edits actually help and applying them correctly.
Anyone running a self-improving agent loop should read this as a routing instruction. Send harness-edit proposals to a cheap model. Reserve frontier capacity for the judgment step, evaluating which proposal to keep, and the apply step, where a bad edit can break the whole loop.
The survey adds a second finding worth acting on. Harness improvements transfer across benchmarks, which means they encode general engineering patterns rather than tricks tuned to one task. A harness improvement discovered while optimizing a coding benchmark isn't wasted if you move to a different task family. That argues for porting harness tuning between projects instead of rebuilding it from scratch every time you stand up a new agent.
Weng doesn't say how large the cost delta is between a Qwen2-32B proposal loop and an all-frontier one, or whether the judgment step itself could be handled by a mid-tier model. Those are the open questions for anyone trying to size the savings before they rewrite a routing layer.
Each link below shares sources, entities, or timing with this story.
Cursor uses Opus / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Cursor uses Opus); both cover Harness, Opus; overlapping topics (harness, model).
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 (benchmark, cost, model).
Opus built by Anthropic / Shared entity: Opus / Shared topic / What happened next
Linked by a graph relationship (Opus built by Anthropic); both cover Opus; overlapping topics (benchmark, cost, model).
Linked by a graph relationship (Opus built by Anthropic); both cover Opus; overlapping topics (cost, engineering, model).
Opus built by Anthropic / Shared entity: Opus / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Opus built by Anthropic); both cover Opus; overlapping topics (cost, model).
Opus built by Anthropic / Shared entity: Opus / Shared topic / Earlier coverage
Linked by a graph relationship (Opus built by Anthropic); both cover Opus; overlapping topics (benchmark, capability, model).
Linked by a graph relationship (Opus built by Anthropic); both cover Opus; overlapping topics (cost, engineering, model).
Lilian Weng works at OpenAI / Shared entity: Opus / Shared topic / Earlier coverage
Linked by a graph relationship (Lilian Weng works at OpenAI); both cover Opus; overlapping topics (benchmark, capability, model).