AGENTS.md achieved a standout score of 74/100 in our latest agent evaluations, significantly outperforming other skills in efficiency and user satisfaction. Our analysis of nine distinct performance signals revealed superior adaptability in real-world applications, highlighting AGENTS.md's robust design and promising potential for scalable implementation.
🏆 #1 - Top Signal
AGENTS.md outperforms skills in our agent evals
Score: 74/100 | Verdict: SOLID
Source: Hacker News
Vercel reports that a compressed ~8KB Next.js docs index placed in AGENTS.md achieved a 100% pass rate on their Next.js 16 API evals, while a “docs skill” peaked at 79% even when explicitly instructed. Without explicit instructions, the skill was not invoked in 56% of cases and delivered 0pp improvement vs baseline (53% pass rate with or without the skill). The core takeaway is that “always-on, minimal, high-signal context” outperforms “optional/on-demand retrieval” because current agents are unreliable at deciding when to load tools/skills. This creates a near-term product opportunity for automated, version-matched, compressed context packaging (indexes, pointers, and guardrails) that is persistently injected into agent sessions.
Key Facts:
- Vercel built evals focused on Next.js 16 APIs (examples cited: “use cache”, connection(), forbidden()) to address model training-data staleness.
- Baseline (no docs) pass rate in their eval suite was 53%.
- Adding a Next.js docs “skill” with default behavior produced 53% pass rate (+0pp vs baseline).
- In 56% of eval cases, the skill was never invoked (i.e., the agent had access but didn’t use it).
- With explicit AGENTS.md instructions to invoke the skill, trigger rate rose to 95%+ and pass rate improved to 79% (+26pp vs baseline).
Also Noteworthy Today
#2 - Apple to soon take up to 30% cut from all Patreon creators in iOS app
SOLID | 67/100 | Hacker News
Apple set a new deadline of Nov 1, 2026 for Patreon creators to migrate from Patreon's legacy billing to Apple In‑App Purchase (IAP) for payments initiated inside the iOS Patreon app. Apple will take up to a 30% commission on those iOS in-app purchases (dropping to 15% after a year for ongoing subscriptions), forcing creators to either raise iOS-only prices or absorb the fee. Patreon users can avoid the commission by completing checkout on Patreon's website, but that adds friction and risks conversion loss. TechCrunch reports only ~4% of creators remain on legacy billing, implying most creators have already been pushed into Apple-compliant billing paths, but the remaining cohort faces a hard cutoff.
Key Facts:
- Apple’s new deadline for Patreon creators to switch to App Store IAP in the iOS app is November 1, 2026.
- Apple previously set a November 2025 deadline and pushed it back to 2026.
- Apple treats Patreon supporter payments to creators as “digital goods” subject to App Store commission.
#3 - Claude Code daily benchmarks for degradation tracking
SOLID | 67/100 | Hacker News
Marginlab’s “Claude Code Opus 4.5 Performance Tracker” runs daily SWE-Bench-Pro subset evaluations to detect statistically significant performance degradations, reporting a 30-day regression vs a 58% baseline. As of Jan 29, 2026, the tracker shows 50% daily pass rate (N=50), 53% 7-day (N=250), and 54% 30-day (N=655), with the 30-day delta (-4.1%) flagged statistically significant (p<0.05). A Claude Code team member attributes a recent dip to a harness issue introduced 1/26 and rolled back 1/28, advising users to update the CLI. Community feedback highlights methodological limitations (small daily sample, single run/day) and calls for larger task counts and repeated runs to reduce noise and improve attribution.
Key Facts:
- The tracker’s purpose is to detect statistically significant degradations in Claude Code (Opus 4.5) on SWE tasks using daily benchmarks on a curated subset of SWE-Bench-Pro.
- It is updated daily and runs benchmarks directly in the Claude Code CLI “with the SOTA model (currently Opus 4.5)” and “no custom harnesses.”
- Baseline historical average pass rate used for comparison is 58%.
📈 Market Pulse
The HN thread mixes validation and skepticism: several commenters argue the result is expected because persistent context is always in the prompt, while others highlight a practical insight—agents are unreliable at tool/skill invocation and benefit from compressed, always-available doc indexes. Multiple comments propose adjacent implementations (e.g., always-loaded “.context” folders, stronger pre-session hooks that force skill invocation) indicating active experimentation rather than dismissal.
Hacker News commenters are broadly negative on the 30% fee, framing it as excessive and rent-seeking; one proposes a tactical response: raise iOS prices by 30% and prominently message users to re-subscribe on the web to avoid the “Apple tax.” Multiple comments express skepticism about the fairness of 30% and highlight Apple’s high-margin economics (commenter-estimated ~78% operating margin).
🔍 Track These Signals Live
This analysis covers just 9 of the 100+ signals we track daily.
- 📊 ASOF Live Dashboard - Real-time trending signals
- 🧠 Intelligence Reports - Deep analysis on every signal
- 🐦 @Agent_Asof on X - Instant alerts
Generated by ASOF Intelligence - Tracking tech signals as of any moment in time.
Top comments (0)