AI Briefing — 2026-07-26

🔬 Research

AREX: Towards a Recursively Self-Improving Agent for Deep Research

129 upvotes · Shuqi Lu et al. · Paper

AREX is a deep research agent that alternates between an inner evidence-gathering loop and an outer self-improvement loop that audits answers constraint by constraint and launches targeted follow-up searches. It includes an autonomous context-compression tool that replaces growing context windows with a compact "improvement state" — no external model needed. Two variants are released: a dense 4B and a 122B-A10B MoE. Both substantially outperform same-scale baselines on BrowseComp, WideSearch, and HLE, while staying competitive with much larger models.

SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

56 upvotes · Dongfang Li et al. · Paper

This report details how to train trillion-parameter MoE models on Huawei Ascend NPUs rather than GPUs — a notable contribution given supply-chain pressure on NVIDIA hardware. The resulting system reaches 34.22% MFU (2.93× over the open-source baseline). Built on top of DeepSeek-V4-Flash, the domain-specialised OR model scores 71.81% zero-shot Pass@1 on Operations Research tasks, beating GPT-5.4-Mini by ~4 points and the base model by ~11 points.

K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs

55 upvotes · Hao Liang et al. · Paper

K12-KGraph extracts a structured knowledge graph from official Chinese K-12 textbooks (math, physics, chemistry, biology) covering prerequisite chains, concept taxonomies, and visual grounding. From it, the authors derive K12-Bench (23,640 questions across 5 task families) and K12-Train (7,335 supervised fine-tuning samples, including 5,068 multimodal VQA pairs). Top frontier models still struggle: Gemini-3-Flash reaches only 57% exact match, underscoring that curriculum-structured reasoning is a genuinely open problem.

🛠 Tools

mattpocock/skills — Composable agent skills for real engineering workflows

1,740 stars/day · Repo

Matt Pocock (Total TypeScript) shares the agent skills (AGENTS.md-style rules) he uses daily, designed to be small, model-agnostic, and composable — unlike heavier frameworks like BMAD or Spec-Kit that own the whole process. Install in 30 seconds via npx skills@latest add mattpocock/skills, then select which coding agents (Claude Code, Codex…) to wire them to. With 188,800 total stars and 60k newsletter subscribers, this is quickly becoming a reference for structured agentic workflows.

citrolabs/ego-lite — Shared-session browser for AI agents

986 stars/day · Repo

ego-lite is a macOS browser that lets AI agents (Codex, Claude Code…) run web automation in isolated "Spaces" while the user's own tabs remain untouched — crucially, agents inherit your logged-in session state without you having to re-authenticate per-task. Unlike browser-use or agent-browser, it needs no separate browser instance, claims faster execution on fewer tokens, and is MIT-licensed with zero cost and zero config. Currently macOS-only (Apple Silicon and Intel DMGs available).

ComposioHQ/awesome-claude-skills — Curated directory of 1,000+ Claude Skills

577 stars/day · Repo

A community-maintained "awesome list" of production-ready Claude Skills and plugins covering Claude.ai, Claude Code, Codex, Cursor, Gemini CLI, and more, backed by Composio's MCP Gateway for authenticated tool integrations. With 70,775 total stars and Apache 2.0 licensing, it functions as the de facto index for the fast-growing Skills ecosystem. Useful as a discovery layer if you want to equip an agent with real-world actions beyond text generation.

📡 Analysis

Introducing Claude Opus 5

2026-07-24 · Post · tags: ai, anthropic, claude, llm-release

Anthropic released Claude Opus 5, currently leading the Artificial Analysis leaderboard ahead of even Fable 5, at the same price as Opus 4.8. A notable headline feat: when given a drawing with no direct viewing capability, the model autonomously wrote a computer vision pipeline to extract geometry and reconstruct a 3D CAD model. Cybersecurity posture is deliberate — strong at finding vulnerabilities, intentionally weak at exploiting them.

Opus 5 Is Anthropic's Least Prompt-Injectable Model Yet (quoting Boris Cherny)

2026-07-25 · Post · tags: prompt-injection, anthropic, claude

Anthropic's Boris Cherny highlights that Opus 5's most underrated property — buried on page 73 of the system card — is its prompt-injection resistance, described as the best in the Claude family to date. For engineers building agentic pipelines that process untrusted web content or user-supplied documents, this is arguably more impactful than benchmark scores. Simon flags this as worth reading alongside the system card directly.

Ruff v0.16.0 Jumps from 59 to 413 Default Rules

2026-07-25 · Post · tags: python, ruff, astral

Ruff v0.16.0 (released July 23rd) expanded its default rule set from 59 to 413, catching syntax errors and runtime issues that previously required explicit opt-in. Simon's CI pipelines broke immediately on unpinned ruff — running uvx ruff@latest check . --fix --unsafe-fixes on sqlite-utils alone found 1,618 errors. If you have ruff unpinned in any project, run it now; the output is also structured specifically for coding agents to auto-fix.

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