Devin’s Fusion harness (now in Desktop and CLI) pairs a frontier “lead” for planning and review with a cheaper “sidekick” for execution—Fable 5.1 + SWE-2 is the recommended combo—matching Claude Code or Codex-level scores at up to ~39% lower cost across coding benchmarks.
Published Oct 28, the report counts about 110 million agent started PRs in September (roughly 6x March), with 90% of agent involved PRs merging within two weeks versus about 86% for the rest; GitHub calls its agent count a floor since it only sees detectable agent activity. New repos with an agent config file averaged eight times the commits of those without one (Octoverse)…
Announced Oct 28, Carbon posts 88.2% on SEC bench Pro, 54.8% on ExploitGym and 71.0% on CWE bench (scored without guardrails) and reportedly doubles Argon's secure code success rate. Access starts with a subset of cyber defenders via Google's Fairwind program, with no date for wider release (Ars Technica)…
Released Oct 28, the open draft spells out how tables, forms, charts and diagrams should be parsed for agents, like [blank] for empty form fields, charts as data tables and diagrams as Mermaid, with example pages for each rule. It's on parseprotocol.com with a Claude skill and ChatGPT reference file to try it, and the repo is MIT code plus CC BY 4.0 text…
Announced Oct 28, T3 Code's fx provider supports Supervised and Auto modes and reuses the models and skills you already set up in fx. Install fx and configure a model provider, then add it in T3 Code under Settings Providers Add provider…
Released Oct 28 under OpenMDW 1.1, the 34B total (about 4.3B active) model transcribes recorded audio up to an hour long in 40 languages, labels up to eight speakers, translates 31 languages to or from English, and takes natural language style instructions. It serves on vLLM or SGLang today, and NVIDIA also shipped the open Granary v2 speech dataset…
Spotted something we missed? Start a thread.