Agent Skills Whitepaper: Kaggle Guide to Procedural Memory for AI Agents
Google/Kaggle Agent Skills whitepaper (May 2026) by Debanshu Das, Tanvi Singhal, and team—progressive disclosure, eval tiers, Skills vs MCP, meta-skills, 40K+ marketplace skills, and the retail case study for enterprise builders.
Agent Skills went from Anthropic experiment to industry default in roughly twelve months. The clearest articulation of why—and how to ship them safely—is the 62-page whitepaper on Kaggle, authored by Tanvi Singhal, Gabriela Hernandez Larios, Debanshu Das, Lavi Nigam, and Smitha Kolam (curated by Shubham Saboo), with the open spec at agentskills.io.
This is not a vendor pitch. It is a builder's manual for two personas:
Builders — use skills in Claude Code, Codex, Cursor, ChatGPT
Developers — write, version, evaluate, and govern skill libraries
The whitepaper names four problems that made multi-agent sprawl the default—and Skills the escape hatch:
1. Context rot
Dump every instruction into one system prompt → LLM performance degrades. Research cited includes Lost in the Middle (Liu et al., TACL 2024) and Context Rot (Chroma, 2025): accuracy falls as input grows—even below the context window limit.
Skills fix: load instructions only when triggered.
2. Procedural memory
LLMs approximate:
Memory type
Analog
Episodic
What happened in this chat
Semantic
Facts in weights / RAG
Procedural
How to execute workflows step-by-step
Skills are the first credible procedural memory primitive for agents—the "runbook the experienced colleague hands you on day one."
Token economics example from the paper: 50 workflows as one prompt = 15,000 tokens/turn. As 50 skills = ~4,000 description tokens + ~2,000 active body ≈ 6,000 total—with 49 bodies on disk.
Anthropic examples cite reductions from ~150K → ~2K active tokens when converting workflows to skills—aligned with Matt Pocock v1.0's 63% savings via the same pattern.
Ball of mud → /improve-codebase-architecture, codebase-design
Contrast with GSD / BMAD / Spec-Kit: Pocock explicitly rejects process frameworks that own the workflow. Skills stay small, composable, adaptable—matching the whitepaper's portability argument.
Enterprise lessons from a solo-author monorepo
Even a one-person library demonstrates governance patterns enterprises need:
---name:cafe-preparationdescription:|
Calculates daily ingredient needs and generates prep sheets.
Use when estimating quantities or shopping lists.
Do NOT use for shift scheduling or accounting.
---
Description = routing algorithm. Front-load trigger keywords; state when NOT to use.
File drop — .agents/skills/ at project root (emerging convention); tool-specific paths vary. Community managers (skillport, openskills) symlink across CLIs.
UI install — Web/enterprise registries; upload folders without terminal.
Programmatic — Google ADK SkillToolset, auto-generates load_skill routing tools.
bash
npx skills install github.com/google/skills
Google launched the official repo at github.com/google/skills at Cloud Next 2026—works across Antigravity, Codex, Claude Code, and compliant agents.
Evaluation: A Skill Without a Test Is a Hope
SkillsBench:19% of tasks performed worse with a skill than without.
Four failure modes
Mode
Symptom
Trigger
Wrong skill fires—or right one never fires
Execution
Triggers but wrong output / tool calls
Token budget
Huge body causes context rot when co-loaded
Regression
New skill breaks routing for existing library
Vercel production data (cited)
56% non-invocation rate for skills expected to fire consistently
Skill stripped of instructions: 58% pass vs 63% with no skill — active harm
Passive AGENTS.md index: 100% pass vs 53% baseline — global context ≠ specialist skills
Target:90% trigger accuracy on descriptions.
Evaluation toolkit (five patterns)
Pattern
Use
Eval-as-unit-test
CI on every change
Golden dataset
Versioned input/output pairs in skill dir
LLM-as-judge
Rubric scoring at scale (swap positions for bias)
Adversarial red-team
Negative boundary + rephrasing tests
Canary/shadow
1% live traffic before action-allowed
Evaluation-Driven Development (EDD): write 3 JSON eval casesbefore drafting SKILL.md:
json
{"case_id":"refund_dup_charge_001","input":"I was charged twice for order #4521 last Tuesday","expected_skill":"refund_processor","expected_tool_calls":[{"tool":"lookup_order","args":{"order_id":"4521"}}],"rubric":["acknowledges duplicate","cites order id"]}
Critical: Never evaluate skills in isolation. Production co-loads 5–15 skills. A 5,000-word body that passes alone may fail in the library.
Read / Draft / Act ladder
Tier
Capability
Gate
Read-Only
Query, describe
90% trigger accuracy
Draft-Only
Produce for human review
20+ golden cases
Action-Allowed
Irreversible ops
Adversarial + pass^k + human sign-off
pass^k: require success on k consecutive runs—not one lucky pass. GPT-4o: 61% pass^1 vs under 25% pass^8 on tau-bench.
Production: Skills Are the Unit of Improvement
Reverse-engineering Claude Code v2.1.88 (Liu, Zhao, Shang, Shen 2026): 98.4% of codebase is operational infrastructure—permissions, compaction, subagents—only 1.6% is the agent loop.
Implication: As models converge on reasoning, deterministic engineering around the model differentiates reliability. The composable unit is the skill.
Improvement style
Cycle time
Context tax
Model swap
Days–weeks
None (weights)
System prompt edit
Minutes–hours
Every turn
Fine-tune
Weeks–months
None
New skill
Hours–days
On-demand only
Google Agents CLI ships seven lifecycle skills (scaffold, build, eval, deploy, publish, observe)—expertise in skills, runtime commoditized.
Context debt: capitalized "ALWAYS DO X" in descriptions → models ignore them. Shift intelligence left into testable scripts/.
Choosing Among 40,000+ Public Skills
Heuristics:
First-party first — google/skills, anthropics/skills, stripe/ai, microsoft/skills
Pin versions — community skills change without notice
Audit like code — supply-chain hygiene; skills run in your context
Source
Trust default
Vendor first-party
Trust; pin version
Org-curated private
Trust within org; PR review
Community
Audit before adopt; pin aggressively
Marketplaces: SkillsMP (1.2M+ indexed), VoltAgent/awesome-agent-skills, addyosmani/agent-skills. For a curated, ranked directory with verification signals and one-click install paths, browse explainx.ai at /skills—we host discoverable skills from many authors alongside first-party repos like Google and Anthropic.
Appendix B Preview: Retail as Canonical Enterprise Case
The whitepaper's retail case study argues: same agent runtime + same MCP APIs ≠ same customer experience. Domain skills are the moat.
Example library:
Skill
Owner
Tier
project-guidance
Trades knowledge
Read-only
materials-list
Pro merchandising
Draft-only
return-policy
Customer service
Read-only (refunds = separate action skill)
100 process variants: one agent + 100 skills (~5K metadata tokens) beats 100 subagents or one 1M-token prompt.
Descriptions are the interface — spend more time here than the body
Skills are dependencies — version, pin, PR review, tests
Right team owns right skill — domain experts, not central AI bottleneck
Runtime is interchangeable — portability is the value
Where to Start Tomorrow
From Appendix A:
Record your best practitioner narrating three workflows (1 hour)
Pick the most repeated; run without a skill; note failures
Draft SKILL.md; write 3 eval cases first (EDD)
Ship read-only tier; iterate description until 90% trigger accuracy
Resist generating fifty skills on day one
Summary
The Kaggle Agent Skills whitepaper makes the case that the format is settled (agentskills.io) while the work around it—evaluation under co-load, meta-skills, DAG composition, enterprise governance—is just beginning.
Skills give agents procedural memory without context rot. They replace many multi-agent deployments with one agent + a versioned library. They fail predictably when descriptions are vague, bodies are bloated, or evals run in isolation.
Start small. Treat skills as code. Measure what you ship. Browse existing skills at explainx.ai /skills before writing your twentieth from scratch.
