Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-05 · high
Weaker models built bigger, more fragmented skill libraries without any real edge over agents that only adapted to context.
Why now: As of August 5, this is the first head-to-head data point on whether skill libraries earn their keep over plain context.
Explicit skill libraries roughly match plain in-context memory across five task domains, per ContinualSkillBench, a new agent benchmark. Each domain runs 100 interconnected subtasks, ordered by rising difficulty and built with real chances to reuse earlier work.
That's a direct test of a common agent design bet, that saving and recalling procedures should beat just re-reading recent context. The benchmark's answer is that the bet mostly doesn't pay off.
Running subtasks in the right order did help performance. But maintaining an explicit skill library performed about the same as an agent that only adapted to what it had just seen.
The paper's reading is that most of the gain looks like an agent getting better at reading its own recent history. It's not building an abstraction it can reuse cleanly.
Weaker models made the pattern worse. They built bigger, more fragmented sets of task-specific skills over the course of the benchmark. None of that extra storage bought them any real edge over agents that just adapted to context.
Each link below shares sources, entities, or timing with this story.
Shared entity: Audit / Same source domain / Shared topic / Earlier coverage / Tension
Both cover Audit; reported by the same outlet (arxiv.org); overlapping topics (context, skill).
Shared entity: Audit / Same source domain / Shared topic / Earlier coverage
Both cover Audit; reported by the same outlet (arxiv.org); overlapping topics (audit, skill).
Shared entity: Audit / Shared topic / Earlier coverage / Tension
Both cover Audit; overlapping topics (audit, context); earlier Audit coverage from 2026-06-28.
Both cover Audit; overlapping topics (audit, context); earlier Audit coverage from 2026-03-02.
Both cover Audit; overlapping topics (audit, skill); earlier Audit coverage from 2026-02-21.
Shared entity: Audit / Shared topic / Earlier coverage
Both cover Audit; overlapping topics (audit, context); earlier Audit coverage from 2026-08-03.
Shared entity: Audit / Same source domain / Earlier coverage
Both cover Audit; reported by the same outlet (arxiv.org); earlier Audit coverage from 2026-08-02.
Both cover Audit; reported by the same outlet (arxiv.org); earlier Audit coverage from 2026-07-30.