AI Briefing — 2026-07-21

🔬 Research

RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources

126 upvotes · Yijia Fan et al. · Paper

RESOURCE2SKILL is a framework that turns multimodal human resources (tutorial videos, repos, articles, reference artifacts) into reusable executable skills stored in a hierarchical "Skill Wiki" combining code, structured text, and visual examples. At inference time, agents retrieve and compose these skills — and can acquire new ones online when coverage is lacking. Across 7 authoring domains it beats no-skill agents by +11.9 pp and wins 26 of 28 model-domain comparisons against strong baselines. The key insight is that videos, code, and articles each carry complementary signals that text-only skill libraries discard.

RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM

126 upvotes · Mikhail Komarov et al. · Paper

RAGU is an open-source, pip-installable GraphRAG engine that separates knowledge-graph construction from retrieval, using DBSCAN deduplication, LLM summarization, and Leiden community detection to avoid the noisy single-pass graphs common in existing systems. Its key bet: graph-construction skills (comprehension, extraction, reasoning over context) scale weakly with model size, so a compact 7B model, Meno-Lite-0.1 (59 downloads on HF so far, Apache-2.0), outperforms Qwen2.5-32B on KG construction (+12.5% relative harmonic mean) at far lower compute. On GraphRAG-Bench (Medical) it achieves evidence recall up to 0.84 vs. ≤0.76 for HippoRAG2, and runs on a single GPU.

SWE-Pruner Pro: The Coder LLM Already Knows What to Prune

59 upvotes · Yuhang Wang et al. · Paper

Rather than attaching an external classifier to prune long contexts, SWE-Pruner Pro exploits the agent's own internal representations: a lightweight head converts them into per-line keep/prune decisions with a length-aware embedding. This cuts up to 39% of prompt+completion tokens across two open-weight backbones and four benchmarks with no task quality loss. On MiMo-V2-Flash it also improves SWE-Bench Verified resolve rate by +3.8% and long-context Oolong accuracy by +2.2 points — a rare case where context compression simultaneously boosts performance.

🛠 Tools

bojieli/ai-agent-book — Open-source book: "AI Agents in Depth: Design Principles & Engineering Practice"

4,434 stars/day · Repo

A fully open-sourced (Apache-2.0) 10-chapter book by Li Bojie, covering AI agent design from first principles to production engineering around the formula Agent = LLM + Context + Tools. It ships with 88 companion experiments (70+ runnable standalone), free PDF/EPUB downloads, and community translations in 5 languages (Chinese, Traditional Chinese, English, Tamil, Vietnamese). With 12,716 total stars and an active push today, it's fast becoming a go-to reference for engineers learning agent architecture.

tirth8205/code-review-graph — Local-first codebase intelligence graph for MCP and CLI

1,833 stars/day · Repo

Builds a persistent graph of your codebase so AI coding tools (via MCP or CLI) retrieve only the relevant context instead of flooding the model with entire repos. The README headline: "Stop burning tokens. Start reviewing smarter." — with benchmarked context reductions on code reviews and large-repo workflows. MIT-licensed, Python 3.10+, MCP-compatible, installable via PyPI; already at 24,057 total stars, pushing twice as fast as the agent book.

diegosouzapw/OmniRoute — Free AI gateway: 268+ providers, 500+ models, auto-fallback

1,107 stars/day · Repo

OmniRoute exposes a single endpoint routing to 268+ providers (50+ free tiers) including Claude, GPT, Gemini, DeepSeek, Kimi K3, and GLM, with quota-aware auto-fallback so your coding agent never hits a hard rate-limit wall. Its RTK+Caveman compression saves 15–95% of tokens, and it supports MCP/A2A, multimodal calls, and direct integration with Claude Code, Cursor, Cline, and Copilot. MIT-licensed, TypeScript, built by 500+ contributors — useful as a cost-reduction and resilience layer for production agent stacks.

📡 Analysis

Reverse-engineering is cheap now

2026-07-20 · Post · tags: reverse-engineering, coding-agents, ai-assisted-programming

Simon Willison notes a pattern: people are now routinely using coding agents to reverse-engineer and automate home devices — tasks previously blocked not by technical difficulty but by poor ROI given the maintenance burden. Agents dramatically lower both the cost of the initial attempt and the psychological cost of future breakage (just regenerate the code). This is a useful mental model for assessing any task where the bottleneck was developer-hour cost rather than feasibility.

Who's Afraid of Chinese Models?

2026-07-20 · Post · tags: qwen, llm-release, ai-in-china, ai-ethics, training-data

Simon highlights Ben Thompson's Stratechery piece arguing the US should legislate training data as fair use and ban distillation bans — leaning into openness rather than fighting the unenforceable. Context: Alibaba just released Qwen 3.8 Max (2.4T parameters, open weights) after initially withholding Qwen 3.7 Max, a reversal Simon attributes partly to Xi Jinping's speech pushing "open source, openness, collaboration." The policy and geopolitical dimension of model releases is becoming as important as the technical one.

Quoting Sam Altman — OpenAI's 2022 open-source strategy exposed

2026-07-20 · Post · tags: sam-altman, openai, ai-ethics

An email from Sam Altman to OpenAI's board dated October 1, 2022, surfaced in the Musk v. Altman lawsuit (2026), shows Altman explicitly framing an open-source GPT-3-level local model release as a competitive moat — "this helps discourage others from releasing similarly-powerful models, and makes it harder for new efforts to get funded." This adds historical context to ongoing debates about OpenAI's open-source posture and directly contradicts narratives around altruistic motivations for openness.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net