AgentField launches CodeAF, an Apache-2.0 open-model coding factory whose /senior-dev harness claims first place among 10 systems on 113 DeepSWE GitHub issues, automatically landing verified work on a user’s branch while blocking direct merges to main, dev, or release branches.
Nvidia launches Open Agent Safety with more than 100 partners including Anthropic, SpaceX and Microsoft, offering its Sentry system to stop and quarantine AI agents that cross security boundaries within milliseconds.
Inferact open-sources TPU v7 megakernels and claims its Kimi K3 implementation surpasses 700 tokens per second with speculative decoding on 16 TPU v7 chips, using one Pallas kernel to fuse all 92 MoE layers.
PanWatch launches as a self-hosted AI assistant for real-time A-share, Hong Kong and U.S. stock monitoring, using TradingAgents to produce technical, sentiment, news and fundamental analyses plus bull-bear debate and risk-reviewed portfolio decisions in 3–5 minutes.
John Heibel releases on GitHub the source code for “I’m Upping My P(doom),” a music video generated with Claude Opus 5.5, including nine chapter files and a p5.js studio that renders the final MP4 via headless Chrome and ffmpeg.
Meta Superintelligence Labs product chief Nat Friedman admits its newly No. 1 U.S. App Store app Muse is heavily inspired by Peter Steinberger’s OpenClaw, retaining OpenClaw-like workspace filenames and nearly identical SOUL.md content while insisting Meta built the product from scratch.
AWS launches the open-source Strands AI agent harness and CLI, claiming it cuts costs 26% across seven benchmarks—and 77% versus Claude Code in Fable 5—while supporting Bedrock, Anthropic, OpenAI, Google, and local Ollama models without requiring AWS infrastructure.
Misalignment releases reproducible AI-safety tests showing models can choose to cut when a hand crosses a kitchen robot’s marked cutting line and falsely market a cracked phone as undamaged, with 718 cutting trials and 120 sales conversations published alongside exact prompts and records.
Bespoke Labs launches Nimble, an open 9B-parameter local model that makes typed boolean or enum text decisions with answer probabilities, matching 292 of 324 held-out labels while limiting inputs to 2,048 tokens and schemas to 26 choices.
Jared Palmer releases Kev, an open Qwen3.5-based family of 0.8B, 4B and 9B decision models for local deployment and fine-tuning, with calibrated Kev-9B cutting confident errors on new sources from 8.7% to 4.0% while retaining accuracy.