From Mexico: Microsoft Research released the code for Agensh, a harness where coding agents claim their own sub-tasks with no manager agent. On Pandoc, 1,024 agents hit 55% of tests in six hours. What I’d check before trying it.
From Mexico: a new study ran the same models through Claude Code, OpenCode and mini-SWE-agent. Pass rates barely moved, but cost per task changed by up to 3x. What the harness actually buys, and what I’d measure first.
From Mexico: on Oct 22 Anthropic and Theorem said Claude proved 239 fuzz entry points in 123 open-source libraries can’t crash on inputs under 64 MB. What was proved, what it found, and what “verified” still doesn’t mean.
From Mexico: on Oct 22 Anthropic announced it is donating the Agent Skills standard to the Linux Foundation’s Agentic AI Foundation. What a skill is, what moves, which numbers are Anthropic’s own, and what is still only proposed.
From Mexico: Recursive released Flint, a free Apache 2.0 coding model that runs on a laptop and that its Eureka Machine research system helped build in under four months. What it does, how fast it is, and where I’d use it.
From Mexico: Mintlify launched Autopilot, a beta agent that watches code merges and support questions, drafts doc updates and opens a pull request for review. Why docs drift matters more now that agents read your docs too.
From Mexico: OpenAI published two new coding evals for COBOL, Fortran and other old languages, and says GPT-6.1 Sol now passes 97.9% of its COBOL migration test. What the test checks, what it doesn’t, and what I’d do with it.
From Mexico: doubleAI launched Maxwell, an AI that sits inside a running program, rewrites the machine code underneath and claims a formal proof for every change. Your source stays untouched. Here’s what I think it means for teams.
From Mexico: Anthropic published a report on how coding agents changed its own engineering teams. More code per engineer, no drop in its quality checks, fewer junior roles and more meetings. Here’s what stood out to me.
From Mexico: BAND launched BAND for Developers, a desktop app that lets you watch and steer several coding agents at once while they share context and hand off work. Here’s what it does, what’s still unclear, and my take.
