Agents Daily

Google Antigravity adds Gemini 4 Hydrogen and Orchestrator model option

Top stories

Google Antigravity adds Gemini 4 Hydrogen and Orchestrator model option

Google Antigravity adds Gemini 4 Hydrogen and Orchestrator model option

  • Google Antigravity is rolling out Gemini 4 Hydrogen and Orchestrator, a model option that switches between Hydrogen and Gemini 4 Carbon within a single task.
  • On Terminal-Bench 4.1, Orchestrator reached near-Carbon success using 24% less quota and finishing 19% faster; Hydrogen alone scores 57% and 51 on the Artificial Analysis index.
  • Hydrogen is available to paid AI Pro and Ultra subscribers, Orchestrator to Ultra, and the Antigravity SDK lets developers wire Hydrogen and Carbon agents together.

Read more at antigravity.google. Also reported by blog.google and aistudio.google.com.

Google releases Gemini 4 Hydrogen for everyday coding and agentic workflows

Google releases Gemini 4 Hydrogen for everyday coding and agentic workflows

  • Google announced Gemini 4 Hydrogen, a model for everyday coding and agentic workflows, available to developers starting November 12.
  • It handles repository coding, debugging, interactive web development and multimodal work across four configurable thinking levels, with tighter instruction following and fewer fabricated claims.
  • Developers can use it via the Gemini API in Google AI Studio and Android Studio, Gemini for business, or Antigravity for paid Pro and Ultra subscribers.

Read more at blog.google

Google Antigravity launches Gemini 4 Carbon and /swarm multi-agent preview

Google Antigravity launches Gemini 4 Carbon and /swarm multi-agent preview

  • Google Antigravity is rolling out Gemini 4 Carbon alongside /swarm, a preview feature where a coordinator runs parallel workers in isolated git worktrees with adversarial verifiers.
  • Carbon ranks #1 on CursorBench 4.0 and Perplexity's WANDR; Google warns swarms consume quota much faster than a standard chat.
  • Google AI Ultra subscribers get Carbon in the App, CLI and IDE extensions, and /swarm is in preview in the App and CLI.

Read more at antigravity.google

In brief

  • Developer says AI-built app prototype lacks the unseen parts of real software. A developer named Bacon writes that a restaurant owner's Claude Code app lacked login, responsive layout, translations and error handling, arguing software's unseen safeguards are the real product. (roll.bybacon.com)
  • Edge0 open-sources Edge0-145B and Edge0-120B local models. Edge0 says its open-source Edge0-145B runs on a laptop with 29GB peak memory and decoding above 80 tokens/s, released with an inference framework and deployment tools. (edge0.ai)
  • Fable52.games arcade collects 32 browser games built by Claude Fable 5.2. The site hosts 32 games built autonomously by Claude Fable 5.2, mostly from one prompt, letting visitors play, read prompts and inspect source; featured is Saltline, an 8-hour FPS build. (fable52.games)
  • Turri open-sources MIT-licensed coding agent built on regenerative software. Turri, built on Pi, has users maintain a system definition of specs, tests and ADRs in ***plain while the agent regenerates code from it and tests check conformance. (github.com)
  • Leeroo open-sources Kapso, an experiment-running and learning engine. Kapso runs experiments, deploys the best solution and banks lessons; its page cites first on RelBench leaderboards, top open-source on MLE-Bench and a 536.07 score at IOAI 2026. (github.com)
  • GitHub issue reports Claude Code 2.1.323 removed subagentStatusLine setting. A GitHub issue says Claude Code 2.1.323 removed subagentStatusLine without replacement or warning, hiding per-subagent context, cost, model and effort, and asks for restoration or prior deprecation. (github.com)
  • Friendly AI Labs announces Hive for running coding agents on remote compute. A Friendly AI Labs co-founder recounts his career, including moving agents to a server farm at Airbnb, and announces Hive, a small beta for running agents from any device. (friendlyai.dev)

Compiled, edited and written with AI. Published by Rapid Review, Inc.

Get Agents Daily by email

Agents Daily is the fast, no-fluff briefing on what actually matters in AI agents and coding agents. Every issue surfaces the key model releases, tools, papers, and security developments so you can stay current without the noise.

  • For developers who build with coding agents
  • Daily, five minutes to read.

You'll get an email to confirm. Unsubscribe any time, from any email. Privacy