INQUIRING LINE

An AI's skill library can sharpen over time — but only if you train a separate curator to decide what stays.

Can skill repositories evolve toward execution-oriented refinement over time?

This explores whether a stored collection of reusable agent skills can, over time, shift from generic accumulated entries toward tighter, action-ready execution logic — and what actually drives that refinement.


This explores whether a stored collection of reusable agent skills can, over time, shift from generic accumulated entries toward tighter, action-ready execution logic. The corpus says yes — but the interesting finding is that the shift doesn't happen on its own. It happens when you separate the thing that *curates* the library from the thing that *uses* it. In SkillOS, a trainable curator sits alongside a frozen executor and, grouped by streams of related tasks, learns to steer the repository away from verbose generic additions and toward actionable execution logic and cross-task meta-strategies — and that curator generalizes across different executor backbones and domains Can a separate trained curator improve skill libraries better than frozen agents?. So execution-oriented refinement is less an emergent property of skill libraries and more a *learned editorial policy* you can train for.

That framing sits against an older, simpler model of skill evolution: refinement through direct environmental feedback. VOYAGER stores executable skills in an embedding-indexed library, composes complex skills from simpler ones, and refines each skill against feedback from the world while an automatic curriculum keeps pushing exploration — enabling lifelong learning without the catastrophic forgetting of weight-update methods Can agents learn new skills without forgetting old ones?. Agent Workflow Memory pushes on granularity instead: it induces reusable sub-task routines finer than whole tasks, abstracts away example-specific values, and compounds them hierarchically, with gains that grow as the gap between training and test widens Can agents learn reusable sub-task routines from past experience?. The direction of travel in all three is the same — from bulky and specific toward compact, composable, and executable.

But there's a bottleneck lurking under 'execution-oriented,' and it's not the skills themselves — it's how you cut the task up to retrieve them. One note finds that standard LLM decomposition reaches only 34% step-level recall, and that simply correcting the *step count* recovers 75% of the achievable gains, shifting the real bottleneck to representation-level reranking rather than vocabulary matching What blocks skill retrieval in task decomposition?. A library can be beautifully refined and still fail if the agent asks it the wrong-sized question.

The deeper reason execution-oriented skills beat offline-authored ones is grounding. MUSE-Autoskill invokes skill creation from *inside* the agent's reasoning loop, so each new skill is born in exact task context with immediate feedback and runtime validation — hitting ~88% task accuracy and transferring to other agents with minimal loss, precisely because it eliminates the mismatch between where a skill was written and where it runs Does creating skills inside the agent loop eliminate mismatches?. This is the same logic that makes tools matter at all: tool-integrated reasoning provably expands what an agent can do, unlocking strategies that are impossible or hopelessly verbose in pure text Do tools actually expand what language models can reason about?. Execution-orientation isn't cosmetic tidying — it's moving skills toward the operations that actually extend capability.

The thing you didn't know you wanted to know: refinement-over-time has a ceiling set by *who is doing the refining*, not just by how much experience accumulates. The Darwin Gödel Machine evolves better agents by keeping an archive of variants and validating them empirically rather than by proof, getting large SWE-bench gains through open-ended self-improvement Can AI systems improve themselves through trial and error? — but a companion finding warns that the *capacity to benefit* from self-edits follows an inverted-U across model tiers: weak models never invoke the refinements, and the strongest models struggle to follow their own instructions faithfully, so mid-tier models improve most Do stronger models always evolve harnesses better?. So skill repositories can absolutely evolve toward execution-oriented refinement — but whether that evolution compounds or stalls depends on decoupling the curator, getting decomposition granularity right, and matching the refiner to a capability band where edits actually stick.


Sources 8 notes

Can a separate trained curator improve skill libraries better than frozen agents?

SkillOS shows that separating a trainable curator from a frozen executor, grouped by task streams, causes skill repositories to shift from generic verbose additions toward actionable execution logic and cross-task meta-strategies. The trained curator generalizes across different executor backbones and domains.

Can agents learn new skills without forgetting old ones?

VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.

Can agents learn reusable sub-task routines from past experience?

Agent Workflow Memory induces sub-task routines at finer granularity than full tasks, abstracts example-specific values, and compounds them hierarchically. This produces 24.6% relative gain on Mind2Web and 51.1% on WebArena, with larger gains as train-test gaps widen.

What blocks skill retrieval in task decomposition?

Standard LLM decomposition reaches only 34% step-level recall, gating retrieval success. Correcting step count recovers 75% of gains in iterative methods, shifting the bottleneck to representation-level reranking rather than vocabulary alignment.

Does creating skills inside the agent loop eliminate mismatches?

MUSE-Autoskill demonstrates that invoking skill creation from within the agent's reasoning loop grounds new skills in exact task context, immediate feedback, and runtime validation. In-loop skills reach 87.94% task accuracy and transfer to other agents with minimal loss, eliminating the situated context problem of offline authoring.

Show all 8 sources
Do tools actually expand what language models can reason about?

Formal proof shows tool-integrated reasoning enables strategies impossible or prohibitively verbose in text alone, expanding both empirical and feasible support. The advantage spans abstract reasoning, not just arithmetic, and Advantage Shaping Policy Optimization stabilizes training without reward distortion.

Can AI systems improve themselves through trial and error?

DGM replaces formal proofs with empirical benchmarking and maintains an evolutionary archive of agent variants, achieving 2.5× improvement on SWE-bench and 2.2× on Polyglot by discovering capabilities like better code editing and context management.

Do stronger models always evolve harnesses better?

Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.

Papers this line draws on 8

The research behind the notes this line reads — ranked by how closely each paper relates.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are a research analyst investigating whether a stored repository of reusable agent skills can evolve toward tighter, execution-ready refinement over time — treat this as still open.

What a curated library found — and when (dated claims, not current truth). These span roughly 2024–2026:
- A trainable curator decoupled from a frozen executor learns an editorial policy that steers a repository away from verbose generic entries toward actionable execution logic and cross-task meta-strategies, generalizing across executor backbones and domains (~2026).
- Inducing reusable sub-task routines finer than whole tasks, abstracted from example-specific values, compounds hierarchically — with gains that grow as the train/test gap widens (~2024).
- Standard LLM decomposition reaches only 34% step-level recall; correcting the step count alone recovers 75% of achievable gains, shifting the real bottleneck to representation-level reranking, not vocabulary matching (~2024).
- Creating skills from inside the agent's runtime reasoning loop, with immediate feedback, hits ~88% task accuracy and transfers to other agents with minimal loss (~2026).
- The capacity to benefit from self-edits follows an inverted-U across model tiers — weak models never invoke refinements, strongest struggle to follow their own edits, so mid-tier models improve most (~2026).

Anchor papers (verify; mind their dates): Agent Workflow Memory (arXiv:2409.07429, 2024); Darwin Gödel Machine (arXiv:2505.22954, 2025); SkillOS (arXiv:2605.06614, 2026); MUSE-Autoskill (arXiv:2605.27366, 2026).

Your task: (1) Re-test each constraint above — judge whether newer models, training, tooling, orchestration (memory, curators, multi-agent), or evaluation have RELAXED or OVERTURNED it; separate the durable question from the perishable limit, cite what resolved it, and say plainly where a constraint still holds. (2) Reconcile the competing threads — decoupled-curator vs. runtime-grounded creation vs. decomposition-granularity — into what actually governs whether refinement compounds; surface superseding work from the last ~6 months. (3) Propose 2 research questions that assume the regime may have moved. Cite arXiv IDs; flag anything you cannot ground in a real paper.