Skip to content

Latest commit

 

History

History
145 lines (102 loc) · 6.67 KB

File metadata and controls

145 lines (102 loc) · 6.67 KB

Resource Evaluation: "AGENTS.md Outperforms Skills in Our Agent Evals"

Date: 2026-01-30 Evaluator: Claude (Opus 4.5) URL: https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals Author: Jude Gao (Vercel) Publication Date: January 27, 2026


Summary

Vercel blog post comparing four documentation strategies for coding agents on 19 Next.js 16 tasks (APIs not in model training data). Finds that a static AGENTS.md docs index achieves 100% pass rate vs skills at 79% (with explicit instructions) or 53% (without, equal to baseline). Core finding: skills were only auto-invoked 56% of the time. A compressed 8KB index (reduced from ~40KB) maintained full performance.

Key metrics:

Configuration Pass Rate
Baseline (no docs) 53%
Skills (default) 53% (+0pp)
Skills (with instructions) 79% (+26pp)
AGENTS.md docs index 100% (+47pp)

Detailed breakdown (Build / Lint / Test):

  • AGENTS.md: 100% / 100% / 100%
  • Skills + instructions: 95% / 100% / 84%
  • Baseline: 84% / 95% / 63%

Evaluation Scoring

Criterion Score Notes
Relevance 3/5 Validates existing CLAUDE.md architecture; indirect (Claude Code ≠ AGENTS.md)
Originality 4/5 First quantified benchmark of eager vs lazy context loading in coding agents
Authority 3/5 Vercel employee, transparent methodology, but conflict of interest (see below)
Accuracy 5/5 All 13 claims fact-checked and verified against source
Actionability 3/5 3 surgical insertions in existing guide sections

Overall Score: 3/5 (Pertinent)


Gap Analysis

Already Covered in Guide

Article Concept Guide Coverage Location
Always-loaded context files CLAUDE.md architecture ultimate-guide.md:4074-4080
CLAUDE.md sizing (4-8KB) Size guideline ultimate-guide.md:3527
Skills as lazy-loaded modules Memory Loading Comparison ultimate-guide.md:4074-4080
skills.sh marketplace Full documentation ultimate-guide.md:5606-5694

What's New

  • 56% invocation rate: First quantified data on skill auto-discovery failure rate
  • 8KB compression benchmark: 5x compression (40KB → 8KB) with zero performance loss
  • Eager vs lazy evidence: First empirical data supporting always-loaded context over on-demand skills for critical instructions

Fact-Check Results

Claim Verified Source
Author: Jude Gao Article byline
Date: January 27, 2026 Article metadata
AGENTS.md 100% pass rate Article table
Skills + instructions: 79% Article table
Baseline: 53% Article table
Compressed index: 8KB Article text
Original size: ~40KB Article ("around 40KB")
Skills invoked 56% of time Article text
Next.js 16 APIs Article (multiple references)
12 APIs listed All enumerated in article
Command: npx @next/codemod@canary agents-md Article code block
Build/Lint/Test breakdown (all configs) Article detailed table
Skills without instructions = baseline Both at 53%

Confidence: High (13/13 claims verified directly in source article + Perplexity cross-check)


Technical Writer Challenge

Agent challenged the evaluation from a documentation perspective:

Key arguments:

  1. Score correct at 3, not 4: The article confirms existing guidance rather than introducing new guidance. A user reading our guide is already doing the right thing.
  2. 56% is the real finding: The headline buries the lead. Skills fail because agents don't discover them, not because skill content is bad. Claude Code already solved this with always-loaded CLAUDE.md.
  3. Sample too small: 19 tasks, ~4 task difference between 100% and 79%. Not statistically robust for broad conclusions.
  4. Self-serving narrative: Vercel operates skills.sh AND authored the AGENTS.md codemod. They're deprecating one positioning and upselling another under the cover of transparency.
  5. Integration scope correct: 3-5 lines, no dedicated section needed.

Score adjustment: None (3/5 confirmed)


System Architect Challenge

Agent challenged the evaluation from an architectural perspective:

Key arguments:

  1. Correct score, wrong reasoning: "Validates our choices" is a non-argument. What saves the 3/5 is the compression benchmark — the only actionable data point.
  2. Missing architectural pattern: The article demonstrates the eager loading vs lazy invocation trade-off. The guide documents this factually (line 4074-4080 table) but never names the pattern or provides empirical evidence for it.
  3. Plan under-specified: Original "3-5 lines, 2 sections" is a placeholder. Correct plan: 3 lines, 3 specific locations (lines 3527, 4082, 5641).
  4. Missing compression technique: The guide says "keep it concise" but never shows how to compress. The 5x compression (40KB → 8KB) with zero loss is actionable guidance the guide lacks.
  5. Decision criteria gap: The guide's loading table (line 4074) lacks a criticality criterion — if instructions are critical, use CLAUDE.md (eager, 100% loaded); if supplementary, skills suffice (lazy, 56% invocation acceptable).

Score adjustment: None (3/5 confirmed), but integration plan upgraded to 3 precise locations.


Conflict of Interest Note

Vercel operates skills.sh (the skills marketplace) and authored the npx @next/codemod@canary agents-md tool evaluated in this article. The article concludes that their own skills.sh platform underperforms compared to AGENTS.md. While this appears intellectually honest (arguing against their own product), Vercel is positioning a different Vercel tool as the replacement. The methodology is transparent and reproducible, so this is noted as context rather than a disqualifier.


Integration Plan

Three surgical insertions in existing sections:

1. CLAUDE.md Sizing (line ~3527)

Add compression benchmark after the size guideline paragraph.

2. Memory Loading Key Insight (line ~4082)

Add warning about skill invocation reliability after the existing key insight.

3. Skills Trade-offs (line ~5652)

Add bullet about invocation reliability.


Decision

Aspect Verdict
Score 3/5 (unanimous across 2 challenger agents)
Action Integrate (3 lines, 3 sections)
Confidence High (13/13 fact-checked + double challenge)
Priority Low (confirms existing architecture)
Transferable insights 56% invocation rate, 8KB compression benchmark, eager vs lazy evidence