Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-16 · high
The preprint pitches an automated behavior map and a step-by-step disclosure method, and picked up 154 HuggingFace Daily upvotes by July 16.
Why now: It picked up 154 HuggingFace Daily upvotes as of July 16, right as more teams ship harnesses that edit their own code on a schedule.
A new paper argues that self-improving AI harnesses hit their real wall at code discovery, not code editing. The Harness Handbook (arXiv 2607.13285) calls this behavior localization, finding every place a given behavior lives before you touch any of it. Miss one site and an agent patches one path while three siblings stay broken, an expensive failure mode once a harness starts rewriting its own code on a schedule.
The fix the authors pitch has two parts. First, an automated map that links each behavior to its source, built from static analysis plus LLM assistance. Second, Behavior-Guided Progressive Disclosure, which walks an agent down a behavioral hierarchy step by step, verifying each candidate code location before it commits to editing there.
The paper reports gains in planning efficiency for behaviors scattered across files and for cross-module interactions, using fewer tokens per fix. It doesn't publish the hard percentages, so there's no way yet to size the improvement against a baseline harness that just greps and hopes. It picked up 154 upvotes on HuggingFace Daily by July 16, which says the harness-building crowd recognized the problem before reading the fix.
The bigger claim buried in here: bigger context windows don't solve this. An agent holding a million tokens still fails if it's reading the wrong files first. Mapping beats memory, and any team running a harness that edits its own code without one is betting on luck.
Each link below shares sources, entities, or timing with this story.
Simon Willison released LLM / Shared entity: LLM / Shared topic / What happened next
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; overlapping topics (agent, behavior, code).
Simon Willison released LLM / Shared entity: Reported / Shared topic / What happened next
Linked by a graph relationship (Simon Willison released LLM); both cover Reported; overlapping topics (agent, harness).
Simon Willison released LLM / Shared entity: LLM / Shared topic / Earlier coverage
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; overlapping topics (agent, argu).
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; overlapping topics (agent, code).
Simon Willison released LLM / Shared entity: LLM / What happened next / Tension
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; picks up the LLM thread on 2026-08-16.
LLM uses OpenAI / Shared entity: LLM / What happened next / Tension
Linked by a graph relationship (LLM uses OpenAI); both cover LLM; picks up the LLM thread on 2026-07-27.
Simon Willison released LLM / Shared entity: LLM / Earlier coverage / Tension
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; earlier LLM coverage from 2026-06-19.
Linked by a graph relationship (Simon Willison released LLM); both cover LLM; earlier LLM coverage from 2026-06-18.