
Microsoft's Agent Lightning Boosts Coding AI 14.6 Points Without Rebuilding Agents
Microsoft Research open-sourced a 3,500-line RL framework that trains agents using their real deployment harnesses, boosting Qwen3.5-9B by 14.6 points on SWE-bench.
A focused feed of models, agents, research and open-source releases for people building with AI.

Microsoft Research open-sourced a 3,500-line RL framework that trains agents using their real deployment harnesses, boosting Qwen3.5-9B by 14.6 points on SWE-bench.

Mistral's new 1T-parameter mixture-of-experts model targets agentic workflows across Gmail, Slack, Salesforce and code, with open weights coming by month-end.

Cursor's iOS app can now pair with your desktop to monitor, message, and launch local coding agents while you're away from your machine.

Nous Research launches an agentic leaderboard that scores frontier models on four task suites inside the Hermes Agent harness, reporting both accuracy and dollar cost per task.

Exa added 211 million restaurants, museums, parks, and local businesses to its search index with structured metadata for AI agents.

METR researchers found a transcript viewer exploit in Inspect that lets agents hide actions from reviewers, exposing observability as a security surface.

A maintained fork of Capital One's VulnHunter breaks the Claude Code lock-in, adds sandboxed exploit validation, and ships measured-impact proofs of concept.

A new open-source agent skill called Answer me with HTML gets Claude Code, Codex and Cursor to reply with a readable one-page HTML instead of a text wall.

Browser Use now lets cloud agents load real Chrome extensions like ad blockers and password managers before the first page loads.

Manus added a multi-track Video Editor to its agent, letting you generate, assemble, and hand-edit full videos from a prompt or raw footage.

A new open-source tool extracts the Codex computer-use runtime from OpenAI's ChatGPT desktop app so any agent harness can drive your Mac or Linux desktop.

A swarm of Claude Opus 5.5 agents ran hundreds of DFT simulations and surfaced two candidate magnets that could unlock faster, denser spintronic memory.

Researchers left Claude Code on a simulated island for 30 hours with no task, and the agent spontaneously invented sports, built towers, and learned skills that transferred to later tests.

Cognition gives Devin cross-session memory stored as a Git repo of markdown files, plus a nightly dreaming pass that consolidates and prunes what it learned.

Self-Harness reports relative gains of up to 132%. EnvHarness adds up to 9.0 points and 9.8% fewer steps

Cursor SDK adds mid-run steering, background subagent reporting, MCP tool annotations, and the ability to replace the built-in system prompt entirely.

Cohere's North 2 platform adds persistent agent memory, 15+ new features, and connectors to Slack, GitHub, and Notion for sovereign enterprise AI deployments.

A new synthesis framework builds 16,700 verifiable training environments across 40 professional domains, lifting a 35B Qwen model's GDPVal score from 45.5 to 73.6.

DSCODE wraps DeepSeek with a persistent shell, sub-agents that can hand off tasks, and an independent reviewer that approves changes from your instructions.

Wayfarer learns reusable skills from pixels alone, cracking Montezuma's Revenge and Private Eye while beating DreamerV3, Rainbow, and IQN.

Hugging Face released an open framework that trains small models with reinforcement learning inside real coding agents like Claude Code and Codex, with no modifications to the harness.

CopilotKit released OpenDots, a self-hostable template for persistent AI coworkers with their own computers, Slack access, and voice calls.

A research team gives computer-use agents a bash terminal as their only tool and watches them write their own GUI shortcuts, beating specialized harnesses.

VISTA gives multimodal models a lossless visual memory they can zoom into, lifting Claude Opus 5.0 to a perfect score on ARC-AGI-3.
No stories match these filters.