Daily Research Brief 2026-07-27

📊 Token usage: estimated from retrieval and writing scale.

Covers the latest AI research, open source and industry moves, updated daily.

Editor’s Note

This weekend’s signal is concentrated: agents are moving wholesale from ‘standalone tools’ to ’engineering infrastructure’. On arXiv, Skill Self-Play and GuardianAgentBench turn ‘co-evolving skills’ and ‘failure mechanisms under adversarial conditions’ into verifiable research questions; on GitHub, mattpocock/skills, DesktopCommanderMCP and OfficeCLI crystallize ‘skill libraries / local machine control / office-file read-write’ into reusable foundations — practitioners should shift effort from ‘prompt-tuning’ to ‘building harnesses + writing skills + adding safety guardrails’. On the industry side, OpenAI’s three-line outage exposed the reliability bill of the agent era.

1. Latest arXiv Papers

1. Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills — https://arxiv.org/abs/2607.22529

2. AgentHOI: Multi-Agent Reasoning for Human-Object-Interaction Video Generation via Implicit Representation Alignment — https://arxiv.org/abs/2607.22241

3. GuardianAgentBench: Where Agents Fail and How to Guard Them — https://arxiv.org/abs/2607.20982

4. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions — https://arxiv.org/abs/2607.20891

5. From Agent Failures to Text Policies: What Works and What Breaks — https://arxiv.org/abs/2607.20668

6. Same Game, Different Story: A Minimal Conservative Strategic Robustness Benchmark for LLM Agents — https://arxiv.org/abs/2607.19670

7. Supra Cognitive Modes: A Routed Architecture for Agent Memory — https://arxiv.org/abs/2607.19096

8. Enhancing Rubric-based RL via Self-Distillation — https://arxiv.org/abs/2607.18082