Google Publishes Two Agent Skill Papers in One Day: WikiSkill and SKILL.state Break New Ground

WikiSkill: Adding a Wiki to Skill Evolution

2.1 Core Idea: Experience → Knowledge → Skill

WikiSkill builds on Karpathy’s LLM‑Wiki concept by inserting a structured knowledge layer (Wiki) between raw experience and executable skill, turning scattered optimization insights into persistent, reusable knowledge.

2.2 Three‑Layer Architecture + Four‑Step Evolution Loop

WikiSkill organizes the agent workspace into three layers and runs a four‑role iteration each round:

Inference Agent : runs rollout on the training set with the current skill (cannot read the Wiki).

Wiki Maintainer : performs root‑cause analysis of trajectories and records failure patterns in Wiki pages.

Skill Proposer : reads the Wiki (ReAct style) and proposes atomic skill updates.

Gating & Rollback : validates candidates on a held‑out set; accepted skills are merged, rejected ones are rolled back, while the Wiki never rolls back, preserving failure records.

This asymmetric design treats skills as roll‑backable hypotheses and the Wiki as immutable knowledge assets.

2.3 Main Experiments: Consistently Winning

Against the strongest baseline, WikiSkill improves average scores by 3.3, 5.1, 10.0, 5.8 and 12.0 points across five benchmarks.

Gemini‑3.5‑Flash jumps from 33.0% to 72.6% on LiveMath and from 50.5% to 76.6% on SpreadSheet.

Baseline methods are unstable: EvoSkill boosts Qwen‑9B on LiveMath but harms Gemma‑4‑31B on the same task.

2.4 Counter‑Intuitive Finding: Skill Evolution Complements Model Scaling

(1) Larger models gain more from skill evolution: Qwen‑4B, 9B and 27B see average lifts of +12.3, +17.5 and +23.9 points respectively; on SpreadSheet the lift reaches +40.9 points for the 27B model.

(2) Skills enable smaller models to “level up”: Qwen‑3.5‑9B with WikiSkill achieves a 47.4% average, surpassing the un‑skilled Qwen‑3.6‑27B (39.4%).

2.5 Skill Transfer Across Models

The skill evolved by Qwen‑3.6‑27B raises Qwen‑3.5‑9B to 70.2% on ALFWorld, higher than the 9B model’s own skill (63.4%).

Even a 4B skill improves Gemma‑4‑31B’s LiveMath to 73.1%.

Transfer can fail: the 4B SpreadSheet skill drags Gemini‑3.5‑Flash down to 18.1% because it contains low‑level Python patches that hinder the larger model.

Conclusion: General procedural knowledge transfers well, while model‑specific workarounds cause negative transfer.

2.7 Case Study: One Wiki‑Guided Skill Evolution

Round 0 : Wiki records a “take‑check‑move” loop; the proposer suggests a goal‑directed action skill, which is rejected but logged.

Round 1 : Using the failure record, the proposer offers a concrete “break‑repetition‑loop” rule (“never put an item back”), which is accepted.

Round 4 : New loop variants appear; the Wiki accumulates evidence and refines the rule to “perform each operation on each object at most once”.

This completes the “audit‑trajectory‑driven evolution” loop.

SKILL.state: Making Skill Execution a Stateful Machine

3.1 Problem: Dialogue‑Style Execution Gets Buried in Its Own History

Prompt length grows linearly with steps, leading to O(T²) token cost.

Stale observations and reasoning linger in context, causing “context poisoning” on long‑horizon tasks.

3.2 Solution: Each Step Sees Only Three Things

P : immutable skill specification.

Σ<t> : structured current execution state (JSON).

ot : latest environment observation.

The model outputs a reasoning trace Rt, a state patch ΔΣ<t>, and an action at. The patch is deterministically verified and merged (Σ<t+1> = Σ<t> ⊕ ΔΣ<t>), after which the reasoning trace is discarded.

3.3 Experiment 1: Long‑Horizon Scaling Saves 16× Tokens and Improves Accuracy

T=100: baseline consumes 1,062,387 tokens; SKILL.state uses 65,408 (16.2× compression) and achieves 0.94 accuracy vs 0.91.

T=200: baseline inflates to 6.1 M tokens; SKILL.state stays at 122 k tokens with 0.94 vs 0.84 accuracy.

Prompt size remains flat at ~1,800 tokens per step.

3.4 Experiments 2‑3: Noise Robustness & State Recovery

Injecting up to 50 noisy telemetry events per step drops standard ReAct from 0.68 to 0.53, while SKILL.state stays ≥0.97 because noise is filtered during state‑patch generation.

In a “silent environment drift” test where an external actor silently changes the world state, baselines hallucinate for 5‑8 steps, whereas SKILL.state recovers with zero extra steps by trusting the current structured state.

3.5 Experiment 4: Real Benchmarks Show Full Lead

InterCode CTF (100 Linux pwn tasks): pass@1 = 54.2%, +7.8 points over the best baseline, with 60.4% token savings; per‑step prompt ≈ 813 tokens.

τ‑Bench Retail: 58.3% success rate, lowest cost.

τ‑Bench Airline: ~2,800 tokens/step (baseline > 11,000), 32.4% success, > 40% token reduction.

3.6 Experiment 5: Structured State Beats Prompt Compression

When all baselines are forced to the same ~1,800‑token budget, ReAct collapses to 0.18 accuracy, LLMLingua to 0.22, while SKILL.state maintains 0.94, demonstrating that the advantage comes from a structured state rather than mere prompt shortening.

https://arxiv.org/pdf/2608.27454
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution
SKILL.state: Scalable Long-Horizon Agent Skills
https://arxiv.org/pdf/2608.26263