Popular News
Latest News
Top Stories
View all
Anthropic’s merges went from 2,500 to 30,000 a day, so it rebuilt source control
From Mexico: Anthropic says its agents now merge over 30,000 changes a day, and git couldn’t keep up. So a seven-person team built Scion, a server-first version control system, in five months. Here’s what changed and what the rest of us can borrow.
A 2B local search agent cut Claude Code’s tokens by a third. Here’s the catch.
From Mexico: LightOn’s ColGrep-Agent-2B runs on your machine and does code search for Claude Code. In their test it cut API tokens by 34% and cost by 40%. The catch: slower answers and an accuracy gap they can’t rule out. Here’s how it works.
Spotify’s AI doubled its code changes. Verification became the bottleneck.
From Mexico: Spotify says AI pushed its monthly merged changes from about 8,100 to 17,000, and it didn’t find AI-written code driving its incidents. What did break was verification, which couldn’t keep up with the pace. Here’s what they changed.
ByteDance’s TRAE puts agent mode and IDE mode under one roof
From Mexico: ByteDance just merged TraeWork and TraeCode into one TRAE app with two modes, one to run agents across projects and one to edit code by hand. Here’s what changed, and why I think agent apps and IDEs are becoming one tool.
Anthropic will scan open source for free. Maintainers choose the firehose.
From Mexico: Anthropic’s models found 29,000+ possible bugs in open source, but people could only review about 6,000. Its new OSS Scanner lets maintainers skip that line and get raw AI reports fast. Here’s what that trade really means.
Agents aren’t making developers careless. They’re making more code.
From Mexico: GitHub’s new data says leaked secrets aren’t rising per push. There’s just way more pushing, a lot of it by agents. Here’s what that means for my repos, and the tiny model GitHub put right at the push to catch passwords.
Reviewing agent code is a trust problem, not a diff problem
From Mexico: JetBrains Research and Lund University say reviewing what a coding agent wrote isn’t about seeing what changed. It’s about knowing where to look hard. Here’s their three-level review idea and what I’d steal from it today.
GitHub is rebuilding Git for repos full of agents
From Mexico: GitHub’s Oct 6 engineering post says agent fleets are breaking the old trade-off in its Git storage, so it’s splitting durable storage from compute and coordinating only the ref update. Here’s what the numbers mean for small teams.
GitLab says coding agents have to earn their autonomy
From Mexico: GitLab published its Security Standard v1.0 on Oct 6, a five-stage ladder that coding agents climb before they can merge on their own, on the same day Transcend pushed /goal flows that stop waiting for a human.
Obelisk 0.42 keeps agent progress in the database, not the process
From Mexico: Obelisk 0.42 (Oct 4) stores coding-agent workflow progress in a database, wraps generated code in nested server/app/deployment grants, and ships a workflow-agent prototype that can deploy and inspect apps after the process dies.
