AI Briefing — May 2, 2026

🔬 Research

Heterogeneous Scientific Foundation Model Collaboration

184 upvotes · Zihao Li, Jiaru Zou, Feihao Fang et al. · Paper

Eywa introduces a framework that lets language models coordinate with specialized scientific foundation models across different data modalities. Instead of forcing everything through language interfaces, it enables predictive models optimized for specific domains to participate in higher-level reasoning. The approach shows promising results in bridging the gap between general-purpose AI agents and domain-specific scientific tools.

GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

87 upvotes · V Team, Wenyi Hong, Xiaotao Gu et al. · Paper

GLM-5V-Turbo treats multimodal perception as a core component of reasoning rather than an add-on to language models. The model integrates visual understanding directly into planning, tool use, and execution workflows, achieving strong performance in multimodal coding and visual tool use. This represents a shift from auxiliary visual interfaces toward natively multimodal agents.

Visual Generation in the New Era: An Evolution from Atomic Mapping to Agentic World Modeling

72 upvotes · Keming Wu, Zuhao Yang, Kaichen Zhang et al. · Paper

This paper argues visual generation should move beyond appearance synthesis toward intelligent systems that understand structure, dynamics, and causality. The authors propose a five-level taxonomy from basic atomic generation to world-modeling systems, highlighting current evaluation gaps that miss structural and temporal failures. The roadmap emphasizes the need for capability-centered rather than perceptual quality assessments.

🛠 Tools

mattpocock/skills — Skills for Real Engineers

3,645★/day · Repo

A collection of composable agent skills designed for real engineering work, not "vibe coding." Created by Matt Pocock and used by 53K+ engineers, these skills focus on practical development workflows like triaging issues, writing specs, and debugging. The 30-second setup integrates with Claude Code, Codex, and other coding agents, emphasizing small, adaptable tools over monolithic process ownership.

warpdotdev/warp — Agentic Development Environment

3,401★/day · Repo

Warp has evolved into an agentic development environment built around the terminal, now with OpenAI as founding sponsor. The platform supports built-in coding agents and external CLI agents like Claude Code and Codex, with 52K+ stars and active development. Features include agent workflow management and a live dashboard showing thousands of AI agents triaging issues and implementing changes.

TauricResearch/TradingAgents — Multi-Agent Financial Trading

2,112★/day · Repo

A comprehensive framework for multi-agent LLM-powered financial trading with 60K+ stars and recent v0.2.4 release. The system coordinates multiple AI agents for trading decisions, risk management, and market analysis. Backed by academic research (arXiv:2412.20138) and an active community, it represents a significant application of agent frameworks to financial markets.

📡 Analysis

Codex CLI 0.128.0 adds /goal

April 30 · Post · tags: coding-agents, system-prompts, agentic-engineering

OpenAI's Codex CLI now implements a "Ralph loop" pattern with the /goal command, allowing the agent to continuously iterate until it evaluates the goal as complete or hits token budget limits. The feature uses automatically injected prompts for goal continuation and budget management, marking another step toward more autonomous coding workflows.

Our evaluation of OpenAI's GPT-5.5 cyber capabilities

April 30 · Post · tags: ai-security-research, gpt

The UK's AI Security Institute evaluated GPT-5.5's cybersecurity capabilities and found it comparable to Claude Mythos for vulnerability detection. Unlike Mythos, GPT-5.5 is generally available, representing a significant expansion in accessible AI security tooling. This follows their previous Mythos evaluation and highlights growing capabilities in AI-assisted security research.

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