Announced Oct 29, the Palo Alto and Singapore public benefit corp is building a "universal verification machine" meant to read code in any language and produce proofs anyone can check, pitching a shift from vibe-coding to "veri-coding" for AI-generated code. It works with the Lean FRO and plans to open-source its foundational libraries, but there's no product to try yet.
Released Oct 29, /goal <objective in the TUI or Desktop has the session's own model review progress after each turn and keep going until it's done, blocked or out of rounds, with /goal pause , /goal resume and /goal clear to steer it. The release also adds opt in scheduled prompts via /schedule and an experimental image generate tool…
Released this week, Droid 1.0 lets the model write a sandboxed JavaScript Script that runs many tool calls in one request (Factory says it cuts cost 15% on Opus 5.5), trims the base prompt 37%, and ships a @factory/droid sdk for building on the same harness. On Factory's own Terminal Bench 4 runs it beat Codex, Claude Code, Cursor, OpenCode and Devin on the same models at lower cost, but those numbers are self reported…
Announced Oct 29, Jules keeps running sessions until Nov 11, then goes read only and shuts down Dec 31, with Google bringing its remote background tasks into Antigravity. You can now export settings, env vars, scheduled tasks, memories, setup scripts and full session history to a local folder, while code and PRs stay in GitHub (The Verge)…
Released this week under MIT, the repo packages the MiniMax Code team's SKILL.md skills for things like subagent split checks, small PRs, clean context diff review, context and tool budget audits, and long session handoffs, and it doubles as a plugin marketplace you can add from Claude Code. Most skills work in any agent that reads SKILL.md , though six engineering skills adapted from Google's practices stay under CC BY 3.0…
Released Oct 25 under Apache 2.0 ( pip install hermit harness ), Hermit runs an RL style loop that evolves an external adapter of instructions, skills, memory and tools while the model and harness stay frozen, then deploys it to Codex, Claude Code or Pi or exports it as a native plugin. Its best reported result is 0.795 on CL Bench database exploration versus a 0.580 in context baseline, but the numbers are self reported, mostly single runs, all on GLM 5.3 Flash…
Spotted something we missed? Start a thread.