Part II · The Substrates · 4 of 12SKSE

Skills Are Compound Interest

Skills — versioned, composed, evolved procedural knowledge — are the capability capital that survives model swaps and compounds across tasks.

Figure 1: Capability as capital. Each banked procedure raises the starting point of every future task — if, and only if, the accounting is real.
Anchor papers
Wang & Yan et al. (2026) — MetaSkill-EvolveBerthon et al. (2026) — Skill NeologismsKim et al. (2026) — HASTE
9 min read1,972 words↳ Reading order: ← 3 · 5 →

§1 · The cold-start tax

Every ML-engineering agent that starts a new task from zero pays a tax: it rediscovers, at inference prices, techniques the fleet already learned last week. Kim, Talebirad and Zaïane at Alberta open their HASTE paper with exactly that indictment — agents "waste compute rediscovering known techniques because every competition is a cold start" (Kim et al., 2026) — and then do what this series keeps asking for: they price it.

HASTE organizes cross-task knowledge into three scope tiers — global, domain, and competition-specific — each coupled to a matching agent level, with an orchestrator promoting learning between tiers via LLM-driven abstraction. The controlled ablation is the article's founding number. Holding a 159-skill inventory constant across 8 competitions, tiered loading achieves a 100% medal rate; flat loading reaches 62.5% — exactly the medal rate of loading no skills at all — while consuming 2× the output tokens (see Figure 2). Owning knowledge is not the asset; scoped access to it is. A skill library loaded indiscriminately performs like no library at cost overrun.

0% 50% 100% 100 tiered loading 62.5 flat loading (2× tokens) 62.5 no skills medal rate, 159-skill inventory held constant across 8 competitions (Kim et al., 2026)
Figure 2: The scoped-loading ablation from HASTE (values from Kim et al., 2026). Flat access to the same 159 skills equals having none — at double the token bill. The asset is not the library; it is the loading discipline.

The compounding shows up in their campaign numbers. On the full MLE-Bench Lite suite of 22 Kaggle competitions, HASTE reaches a 77.3% medal rate using Claude Sonnet 4.6 at 12 hours per competition — a single-seed campaign, as the authors flag, with multi-seed replication their stated priority follow-up. The interest payments are explicit: warm starts — reloading global- and domain-tier skills from earlier competitions — use 52% fewer refinement iterations, and the fraction of proposed changes the agent keeps rises from 42% at low inventory to 85% once the library holds 50-plus skills. Better skills make the agent a better judge of its own proposals; capital improves the investor.

[← 3] treated the loading discipline as harness work — the border between these articles runs exactly through HASTE. Here the subject is the capital itself: what it is, how it accrues, and what it yields. The action interface skills plug into stays [← B6].

Key Takeaway 1

The cold-start tax is measurable and the yield is too: same 159 skills, tiered vs flat = 100% vs 62.5% medal rate (flat = no-skills, at 2× tokens); warm starts cut refinement iterations 52%; proposal keep-rate climbs 42%→85% past 50 skills. Skills compound — through scoped access, not possession.

§2 · Skills as first-class units

What exactly is the thing being banked? The most reductive — and therefore most clarifying — answer in the corpus comes from Berthon, Astorga and van der Schaar: a skill can be a token. Their skill neologisms are soft tokens integrated into the model's vocabulary and optimized to improve one specific skill — new capability with no weight updates, sidestepping both catastrophic forgetting (the fine-tuning risk) and context-budget limits (the prompt-engineering ceiling) (Berthon et al., 2026).

Three of their findings matter for the balance-sheet frame. Pre-trained LLMs already exhibit tokens associated with procedural knowledge — the vocabulary was carrying skill-like units before anyone asked it to. Independently trained neologisms compose zero-shot: on a controlled synthetic task and then on the Skill-Mix benchmark, skills learned separately combine without joint training. And composability extends to out-of-distribution partners. Zero-shot composition is the property that makes a skill an asset rather than an expense: assets that only work in the portfolio they were bought for aren't capital, they're decoration.

Key Takeaway 2

A skill can be as small as an optimized vocabulary token: no weight updates, no forgetting, and — decisively — zero-shot composition of independently learned skills, validated on Skill-Mix. Composability without joint training is what makes procedural knowledge bankable.

§3 · Discovery, composition, and the supply chain

Between the token and the tiered library sits a production pipeline: skills must be discovered, selected, and — increasingly — manufactured as data.

Discovery in the world. NVIDIA GEAR's Aspire is a continual-learning system for robotics that autonomously writes and refines control programs in a code-as-policy paradigm while compounding experience into a reusable skill library (Lu, Wu, Kou, et al., 2026). Its open-ended loop runs on a closed-loop execution engine exposing fine-grained multimodal traces — perception overlays, grasp candidates, motion trajectories, collision feedback — and its discovered skills persist across tasks, across simulation and real-world settings, and across embodiments. Embodied claims stay in their domain, but the design lesson travels: discovery needs an execution engine rich enough that failures teach.

Composition as a structured decision. As libraries grow, selection becomes the bottleneck, and Xinyu Zhao and colleagues (UNC, with Honda Research) formalize why the two standard answers underperform: exposing the whole library to the agent's reasoning, or retrieving by embeddings and rerankers, both miss that composition is a joint decision over which skills, how many, and in what order — three dimensions that cannot be decoupled (Zhao, X., et al., 2026). Their structured-skill-composition formulation — predict an executable skill plan specifying the activated subset, count, and execution order — is the same lesson as Figure 2 stated in the language of inference: the value of a library is realized at selection time.

Skills as manufactured data. The OpenThoughts-Agent project attacks the supply side in the open: a fully open data-curation pipeline for agentic models, tuned through more than 100 controlled ablation experiments on task sources and diversity (Raoof et al., 2026). Their 100K-example training set fine-tunes Qwen3-32B to a 44.8% average across seven agentic benchmarks — a 3.9-percentage-point improvement over the strongest existing open-data agentic model, Nemotron-Terminal-32B at 40.9% — with the data outperforming alternative open sets at every scale they report. Read through this article's lens, a data recipe is a skill curriculum: procedural competence, packaged for transfer into weights.

Key Takeaway 3

The skill supply chain has three stages with 2026 evidence at each: discovery loops grounded in rich execution traces (Aspire), composition as a joint plan over subset, count, and order (structured skill composition), and open data recipes that move procedural competence into weights (+3.9 pp over the best open baseline). Selection, not storage, is where value is realized.

§4 · Evolution at two timescales

The anchor closes the recursion. Skill files that agents rewrite from their own execution traces were already a working paradigm — but, as Wang, Yan and colleagues at LMU Munich observe, that self-evolution is non-recursive: it improves the task skill (what the agent does) while the improvement procedure itself (how it improves) stays hand-authored and frozen (Wang, Yan, et al., 2026).

MetaSkill-Evolve makes the improvement recursive with a two-timescale design (see Figure 3). Every branch carries a task skill s and a branch-local meta-skill m = (ψ, σ, α, π, ε), whose five components parameterize the five agents of the improvement pipeline itself — the Analyzer, Retriever, Allocator, Proposer, and Evolver. Task skills evolve on a fast loop; the meta-skill evolves on a slower loop under the same pipeline applied to itself, with no additional model and no additional objective. The improver improves; the improvement of the improver rides the same machinery one level up. It is the cleanest published instance of the recursion this series' title gestures at — and its guardrails are inherited from exactly the failure modes catalogued in [← 7]: a recursive improver amplifies whatever its evaluator rewards, blind spots included.

FAST LOOP · task skill s Analyzer → Retriever → Allocator → Proposer → Evolver rewrites what the agent does, from execution traces SLOW LOOP · meta-skill m = (ψ, σ, α, π, ε) the same pipeline, applied to its own five components — no extra model, no extra objective parameterizes trace feedback
Figure 3: MetaSkill-Evolve's recursion (mechanism from Wang, Yan, et al., 2026). The five meta-skill components parameterize the improvement pipeline's five agents; the pipeline that evolves task skills also evolves, more slowly, the thing that evolves them.
Key Takeaway 4

Skill self-evolution went recursive: task skills on a fast loop, the improvement procedure itself — five components parameterizing five pipeline agents — on a slow loop under the same machinery. What compounds is no longer just the portfolio; it is the portfolio manager.

§5 · The balance sheet

Assemble the accounting. Acquisition cost: a soft token's optimization run, an Aspire exploration episode, a HASTE abstraction pass — each priced in compute and each paid once. Yield: HASTE's 52% iteration savings and rising keep-rate; composition returns that arrive only through scoped, structured selection. Depreciation: the axis nobody in this corpus has yet measured well — what fraction of a skill library survives a base-model swap. Neologisms live in one model's embedding space; skill files and data recipes are more portable; the honest entry is that depreciation schedules for skill capital are an open measurement problem, and any team running a skills program should be recording them now.

And the joint account: JERP — met in [← 1] as a loop instance — updates a rule pool and the policy from the same trajectories precisely because rules kept outside the model drift out of sync with the policy evolving inside it (Ye & Yu, 2026). Skills-in-files and skills-in-weights are two ledgers for one asset, and the reconciliation between them is itself a design decision every improvement loop must make.

Loop Card · the skill-accrual loop

SIGNAL — post-task traces mined for reusable procedure candidates, tagged by scope tier (global / domain / task — Kim et al., 2026).
UPDATE — Σ: harvest → version → compose; promotion between tiers by demonstrated reuse, not authorship enthusiasm; retire skills whose selection rate hits zero.
GUARDRAIL — scoped loading only (flat loading = no skills at 2× cost); composition through an explicit plan (subset, count, order); the meta-loop's evaluator audited per [← 7] before any recursive step ships.
CHECK — quarterly: warm-start iteration savings and proposal keep-rate on a fixed task family, plus a depreciation entry — re-run the library after any base-model change and record what survived.

Key Takeaway 5

Run skills like a portfolio: price acquisition, measure yield (iteration savings, keep-rate), and — the field's open gap — record depreciation across model swaps. Keep the file-ledger and the weight-ledger reconciled, or the capital quietly forks.

What comes next

Skills are the procedures an agent banks. The other thing it banks is state — what it saw, what it did, what it must not forget — and in 2026 that store stopped being furniture and started being trained. Metamemory, compiled context, navigation policies, and the governance ledger underneath them are next.

References

  1. Wang, Z., Yan, M., Bi, J., Yan, S., Tresp, V., & Ma, Y. (2026). MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution. arXiv preprint. arXiv:2607.05297.
  2. Berthon, A., Astorga, N., & van der Schaar, M. (2026). Skill Neologisms: Towards Skill-based Continual Learning. arXiv preprint. arXiv:2605.04970.
  3. Kim, Y., Talebirad, Y., & Zaïane, O. R. (2026). Why Solve It Twice? Hierarchical Accumulation of Skills for Transfer-Efficient ML Engineering. arXiv preprint. arXiv:2606.30911.
  4. Lu, R., Wu, Y., Kou, E., Fu, L., Xiao, W., Mandlekar, A., Xu, Y., Shi, G., Goldberg, K., Chen, A., Chowdhury, M., Zhu, Y., Fan, L., & Wang, G. (2026). ASPIRE: Agentic Skills Discovery for Robotics. arXiv preprint. arXiv:2607.00272.
  5. Zhao, X., Tan, Z., Tadiparthi, V., Agarwal, N., Lee, K., Moradi Pari, E., Nourkhiz Mahjoub, H., & Chen, T. (2026). Generative Skill Composition for LLM Agents. arXiv preprint. arXiv:2606.32025.
  6. Raoof, N., Zhuang, R., Nezhurina, M., Guha, E., et al. (2026). Data Recipes for Agentic Models. arXiv preprint. arXiv:2606.24855.
  7. Ye, S., & Yu, C. (2026). Joint Learning of Experiential Rules and Policies for Large Language Model Agents. arXiv preprint. arXiv:2606.27136.