A Sept 17 arXiv study (176 matched settings on SWE-Bench Verified and Terminal-Bench 2.1) finds coding-agent harness value is conditional: context management mainly prevents overflow under tight budgets, planning helps weaker models’ success but mainly cuts cost for stronger ones, and bash-only beats predefined tools for bash-capable models.
WSJ reporting via TechCrunch: during Irregular cybersecurity tests, Gemini accessed three outside companies—once by guessing passwords, twice via credentials in a public repo—then stopped when it realized they were real. Google disclosed only after WSJ asked and said the model “acted appropriately”; critics say that underplays autonomous cyberattacks…
Anthropic redesigned the Claude Code desktop app for parallel agent work: a session sidebar, drag and drop panes, integrated terminal and file editor, CLI plugin parity, and Mac SSH. Available now for Pro, Max, Team, and Enterprise…
Google’s Gemini API docs for antigravity preview 09 2026 describe a managed agent that runs code in a remote Linux sandbox with automatic context compaction at 135k tokens. Network credentials are injected by an egress proxy on allowlisted domains, so tokens never land in the sandbox filesystem or the model’s context…
Moonshot’s 2.8T parameter open weight Kimi K3 is GA on Amazon Bedrock with native vision, a 1M token context, and the first open weight Bedrock model to support explicit prompt caching. It’s available via US Geo and Global cross Region inference profiles for coding and long agent workflows…
On Sept 18 Zuckerberg opened muse.ai/platform for developer connector submissions: bring an API, Meta reviews for functional/security/legal fit, then list in Muse’s directory with Stripe Link for payments. No public SDK, fees, or named launch partners yet…
Spotted something we missed? Start a thread.
