AI Briefing — 2026-06-25

🔬 Research

Qwen-AgentWorld: Language World Models for General Agents

103 upvotes · Zuo, Xiao, Sheng et al. (Qwen/Alibaba) · Paper

Qwen introduces two language world models — Qwen-AgentWorld-35B-A3B and -397B-A17B — trained on 10M+ real-world interaction trajectories across 7 domains, capable of simulating agentic environments via long chain-of-thought reasoning. The three-stage pipeline (CPT → SFT → RL with hybrid rubric-and-rule rewards) yields models that significantly outperform frontier models on the new AgentWorldBench. Beyond simulation, using world-model training as a warm-up for downstream agents improves performance across 7 agentic benchmarks, suggesting language world models can serve as both environment simulators and agent pre-trainers. Code is available at QwenLM/Qwen-AgentWorld.

Are We Ready For An Agent-Native Memory System?

60 upvotes · Zhou, Zhou, Han et al. · Paper

This paper evaluates 12 agent memory systems across 5 benchmarks (11 datasets) through a data-management lens, decomposing memory into four modules: representation/storage, extraction, retrieval/routing, and maintenance. The key finding: no single architecture dominates all scenarios — effectiveness depends on workload alignment — and localized maintenance is more cost-efficient than global reorganization. For engineers building memory-augmented agents, this is the most systematic comparison to date, with actionable guidance on trade-offs between retrieval precision, update correctness, and long-horizon stability. Code at OpenDataBox/MemoryData.

NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?

53 upvotes · Wang, Cheng, Zuo et al. · Paper

NatureBench is a 90-task benchmark derived from real peer-reviewed Nature-family papers, with each task running in an isolated containerized environment (NatureGym) to ensure reproducibility. Under a strict no-web-search protocol, the best tested agent exceeds SOTA on only 17.8% of tasks. Failure analysis shows agents mostly reframe problems as supervised prediction tasks rather than engaging in genuine scientific reasoning, with wrong method choice (not task misunderstanding) being the dominant failure mode — a sobering signal about current coding agent capabilities on real research.

🛠 Tools

calesthio/OpenMontage — Agentic open-source video production system

3,719 stars/day · 21k total stars · Repo

OpenMontage exposes 12 pipelines, 52 tools, and 500+ agent skills that let an AI coding assistant (e.g. Claude, Cursor) orchestrate a full video production workflow: research, scripting, asset generation, editing, and final composition — all from a plain-language prompt. Licensed under AGPLv3, it reached #1 on GitHub Trending the day of release. The "paste a video you love" feature lets users import an existing video as a style reference, making it practically accessible for rapid content creation.

ZhuLinsen/daily_stock_analysis — LLM-powered multi-market stock analysis

1,468 stars/day · Repo

A Python system that aggregates multi-source market data and real-time news, runs LLM-driven analysis across multiple markets, and delivers automated notifications with a decision dashboard. Notably designed for zero-cost scheduled runs, making it accessible without cloud spend. Useful as a reference architecture for combining financial data pipelines with LLM reasoning in a production-ready setup.

NousResearch/hermes-agent — Self-improving agent with persistent memory

1,178 stars/day · 202k total stars · Repo

Hermes Agent from Nous Research is a self-improving agent featuring a built-in learning loop: it synthesizes reusable skills from experience, improves them during use, searches its own past conversations, and builds a persistent user model across sessions. It supports any model via OpenRouter (200+ models) or Nous Portal, runs on anything from a $5 VPS to serverless infrastructure, and is accessible remotely via Telegram. MIT-licensed with 202k stars, it's one of the most starred agent frameworks on GitHub right now.

📡 Analysis

simonw/browser-compat-db

24 juin 2026 · Post · tags: github, mozilla, mdn, datasette-lite, ai-assisted-programming, model-context-protocol

Simon Willison converted Mozilla's entire mdn/browser-compat-data repo into a ~66MB SQLite database, inspired by the new MDN MCP server. The conversion script was generated with Claude Code (Opus 4.8), and a GitHub Actions workflow (built with Codex Desktop / GPT-5.5) force-pushes the database to an orphan branch to benefit from GitHub's open CORS CDN headers. The result is explorable directly in-browser via Datasette Lite — a neat pattern for making static structured datasets queryable without a backend.

Quoting Tom MacWright — "Accidental Anonymity"

24 juin 2026 · Post · tags: careers, ai, ai-misuse

Tom MacWright observes a new pattern in hiring: job applications where every layer — resume, portfolio site, GitHub projects, commit messages — is LLM-generated, resulting in candidates who are effectively anonymous. The problem isn't AI assistance per se but the erasure of any signal about the actual person. For engineers and hiring managers alike, this is a timely reminder that full-stack LLM polish can actively hurt a candidate's chances.

datasette 1.0a35

23 juin 2026 · Post · tags: datasette

The latest Datasette alpha ships a new "Create table" UI backed by a JSON API (/<db>/-/create) supporting columns, primary keys, constraints, and foreign keys, plus a full "Alter table" interface for renaming/reordering/dropping columns and changing types. Template context variables are now formally documented as a stable API until Datasette 2.0, generated directly from dataclass definitions in the view code. This release substantially narrows the gap between Datasette and a full database management UI.

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