#coding-agents

A phone agent startup says Mercury cut its worst-case response from minutes to one second. Inception Labs' new diffusion model is built around that kind of number.

An arXiv study clocked an LLM repair loop damaging correct programs at 0.261 while fixing buggy ones at 0.023, then found the internal direction driving it.

GitHub's Project HydraFusion assembles a multi-model plan for every Copilot coding request, trading single-model simplicity for sharply lower cost.

OpenAI says agents now run 3.1 workdays of effort per human workday in its research org, but most long successful tasks still need human intervention.

Cognition says SWE-2 scores 50.0% on FrontierCode 1.1 Main, a point behind Fable 5.1, at 64% lower cost — built on an open Chinese base model.

Investors just paid twice as much for Cognition as they did in May, and the company is not yet half as large as the competitor that sold out four months ago. Th...

A scan of 6,214 domains found 120 llms.txt files pointing coding agents at unregistered packages, and one slot already held live malware.

Laude Institute's Headlong is an open source agent microharness under 10,000 lines of Bash, built so agents keep thinking between conversations.

MCP's maintainers published five priority areas covering long-running agent work, transport unification, agent identity and the context cost of tool sprawl.

Prime Intellect ran 153 autonomous jobs across 18 frontier models on the nanoGPT speedrun, and the token accounting tells a different story than the ranking.

Slack Code launches with Claude, Devin, GitHub Copilot and Vercel. Agents open the channels, humans review or stop them, and archives serve as audit logs.

DeepReinforce's Ornith-1.5 posts 86.1 on Terminal-Bench 2.1 against Claude Opus 4.8's 85.0, and ships a 9B variant small enough to run on a phone.