#most-interesting-this-week 1 min read
The most interesting stuff this week
Agents reviewing code, swarms picking leaders by list order, and a 1-trillion-parameter open model.
Code review
- How the Claude Code team does code review. PRs doubled while reverts fell 40%.
- AI agent fixed the bug, but what else did it change? Agent fixes caused 21 regressions in 130 patches.
- Catching Bugs Before They Land. AI-written tests caught 74 serious privacy bugs.
Agent swarms
- The surprising leadership dynamics in agent swarms. Agents in swarms mostly ignore each other's orders.
- We ran AI agent swarms in Minecraft and Mindustry. Swarms pick leaders by list order.
Running agents
- GitLab Orbit is GA: Higher agent accuracy, lower token cost. A code graph beat RAG for AI review.
- TokenEconomizer. Remembering repo facts cuts agent costs up to 31%.
- OpenShell. Sandboxed agents never see your real credentials.
- Making every harness durable. Agents can now survive crashes mid-task.
Models
- Le Chonk: Built for Coders, Trusted by Defenders. Mistral released a trillion-parameter model you can download.
- Introducing Flint. A strong coding model that runs on laptops.
- Apple's 20B On-Device Model Uses Just 2GB of Memory. It uses far less power, but solves fewer problems.