AI Briefing — May 10, 2026
🔬 Research
Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction
70 upvotes · Li et al. · Paper
This work challenges conventional retrieval systems by proposing Direct Corpus Interaction (DCI), where AI agents search raw corpora using terminal tools like grep and shell commands instead of vector databases. DCI outperforms traditional sparse/dense retrieval on IR benchmarks and agentic tasks, requiring no offline indexing while adapting naturally to evolving local data. The approach suggests that as language models improve, the interface resolution matters as much as reasoning ability for effective search.
Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning
63 upvotes · Shi et al. · Paper
Skill1 introduces a unified framework that co-evolves three critical agent capabilities—skill selection, utilization, and distillation—using a single task-outcome reward signal. Unlike previous methods that optimize these abilities separately, this approach trains one policy to search skill libraries, apply skills, and distill new ones from experience. Results on ALFWorld and WebShop show superior performance compared to existing skill-based and RL baselines.
MiniCPM-o 4.5: Towards Real-Time Full-Duplex Omni-Modal Interaction
60 upvotes · Cui et al. · Paper
MiniCPM-o 4.5 achieves simultaneous seeing, listening, and speaking through its Omni-Flow framework, enabling proactive behaviors like issuing reminders based on continuous scene understanding. At 9B parameters, it approaches Gemini 2.5 Flash performance while running real-time full-duplex interaction on edge devices with under 12GB RAM. The model represents a shift from turn-based to time-aligned multimodal interaction, with strong adoption evidenced by 286K+ downloads of the MiniCPM-o family on HuggingFace.
🛠 Tools
anthropics/financial-services — Claude for Financial Services
3,281 stars/day · Repo
Anthropic has open-sourced reference agents, skills, and data connectors for financial workflows including investment banking, equity research, and wealth management. The toolkit works both as Claude Cowork plugins and through the Claude Managed Agents API, featuring agents like Pitch Agent and Market Researcher that draft analyst work products for human review. With 17.8K total stars and active development, it provides production-ready templates for FSI workflows while maintaining strict disclaimers about human oversight requirements.
addyosmani/agent-skills — Production-grade engineering skills for AI coding agents
3,009 stars/day · Repo
A curated collection of engineering skills and best practices specifically designed for AI coding agents, implemented in Shell scripts. The repository aims to bridge the gap between AI capabilities and production-grade software engineering practices. Given the rapid star growth and focus on production readiness, this appears to be addressing real pain points in deploying AI agents for code generation and software development workflows.
datawhalechina/hello-agents — 📚 《从零开始构建智能体》
1,197 stars/day · Repo
A comprehensive Chinese tutorial for building AI agents from scratch, covering fundamental principles and practical implementations. Published by Datawhale China, this educational resource addresses the growing demand for agent development knowledge in the Chinese AI community. The steady daily growth suggests strong interest in accessible agent development education, particularly in non-English markets.
📡 Analysis
Quoting Luke Curley
May 9 · Post · tags: webrtc, openai
Luke Curley (formerly Discord) highlights a fundamental mismatch between WebRTC's audio optimization and AI voice interactions. WebRTC aggressively drops packets to maintain low latency for real-time conversations, but this degrades expensive LLM prompts where users would prefer accuracy over speed. This technical constraint reveals why OpenAI's voice features might feel inconsistent—the underlying protocol prioritizes conference call dynamics over AI interaction patterns.
Using Claude Code: The Unreasonable Effectiveness of HTML
May 8 · Post · tags: html, security, markdown, ai, prompt-engineering, generative-ai, llms, llm, claude-code
Thariq Shihipar from Anthropic's Claude Code team advocates for requesting HTML output from LLMs instead of Markdown, enabling richer explanations with SVG diagrams, interactive widgets, and enhanced navigation. Simon Willison demonstrates this with a security exploit analysis, showing how HTML's capabilities can make complex technical content more digestible. This shift from token-efficient Markdown to feature-rich HTML reflects the evolution from token-constrained to capability-focused AI interactions.
Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net