AI Agent周报：Gemini 3.5 Flash、Cohere开源等

In today’s issue:

• Google ships Gemini 3.5 Flash for agents
• Antigravity 2.0 becomes a full agent platform
• OpenAI ships Appshots and /goal in Codex
• Cohere open-sources Command A+ on Apache 2.0
• Qwen3.7-Max runs agents for 35 hours straight
• NVIDIA verifies agent skills
• Cursor Composer 2.5 sharpens coding agents
• Anthropic acquires Stainless for SDK tooling
• Browserbase opens Browse.sh skills catalog
• Gemini Omni unifies create-anything model
• OpenAI cracks an 80-year Erdős problem
• Compiling agent workflows into model weights
• PEEK orientation cache for long-context agents
• SaaS-Bench exposes computer-use agent ceiling

And all the top AI dev news, papers, and tools.

Top Stories

Gemini 3.5 Flash and Managed Agents Land

Google opened I/O 2026 with Gemini 3.5 Flash, a frontier model tuned explicitly for agents and coding, alongside Managed Agents in the Gemini API that ship an isolated execution environment with every request.

• Agentic benchmarks: Gemini 3.5 Flash posts 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, and 1656 Elo on GDPval-AA, outperforming Gemini 3.1 Pro on long-horizon coding and tool-use tasks at 4x faster output.
• Managed Agents preview: A single Gemini API call spins up an agent that reasons, uses tools, and executes code in an ephemeral Linux sandbox managed by Google, with AGENTS.md and SKILL.md as versionable config.
• Where it ships: Available in Google AI Studio, Android Studio, Antigravity, Gemini Enterprise Agent Platform, the Gemini app, and AI Mode in Search, with 3.5 Pro slated for next month.
• Why it matters: Flash is now the cost-optimized agent default at Google scale, and Managed Agents removes the build-your-own-sandbox tax that has kept many teams on third-party runtimes.

Blog | Managed Agents

Read more