CodeScene’s agents (mostly Claude Code + Opus) refactored a 300K-line Street Fighter III decompilation in three weeks for ~$4k—2,903 commits, Code Health 5.6→10.0—guided by a CodeHealth MCP score and frame-by-frame replay checks; they also accumulated a 22-recipe refactoring playbook.
Anthropic finds Z.ai’s freely downloadable GLM 5.3 builds end to end exploits at rates close to Claude Mythos Preview (50/410 vs 56/410 on ExploitBench), while simple jailbreaks and weight abliteration bypass its safeguards 64–100% of the time—unlike safeguarded Claude models in the same tests…
OpenClaw shipped OCE, a free MIT licensed, self hostable control plane (Docker/Kubernetes) for multi tenant agent deploy with permissions, sandboxing, and audit—framed as “Kubernetes for agents.” The project started at OpenAI and now lives under the OpenClaw Foundation with Red Hat and Nvidia; OpenAI and Red Hat are already piloting it (still pre 1.0, aimed at internal pilots)…
MLC’s TIRx Harness pairs a thin PTX level compiler with a kernel zoo, sync/race diagnostics, and a remote benchmark server so coding agents can iterate on GPU kernels without measurement noise. On Blackwell, agents hit family geo mean speedups from 1.33× to 6.84× vs baselines (KDA forward 2.94× over FlashKDA; backward 6.84× over FLA)…
At DevDay (Sept 29), OpenAI opened a limited preview of Decisions API on Luna: you supply a question, finite answer set, and text/image context; it returns one choice with a confidence score in about 150ms—aimed at classification, routing, and picking an agent’s next step. Broad rollout is promised in the coming days; pricing and full docs are not public yet…
Hugging Face’s Serge overnight picks persistent Transformers GPU CI failures, reproduces them, writes a patch, verifies both trees on GPU, and only then opens a maintainer PR—29 merges in 80 days ( $43 inference per merged PR on Kimi K 2.7 Code)…
Spotted something we missed? Start a thread.
