Google's open intelligent agent underlying infrastructure: supports single-call cloud Linux sandbox, handwritten Markdown to start AI.

AIMPACT News. On May 20 (UTC+8), according to Beating monitoring, agent development is moving on from cumbersome server operations and sandbox setup, directly shifting to cloud-based transient calls that define configuration files. At I/O 2026, Google officially launched a preview of Gemini API Managed Agents, opening its underlying agent infrastructure to developers. With just a single API call, developers can instantly spin up an isolated, temporary Linux environment hosted by Google in the cloud, running an Antigravity agent driven by the next-generation mid-tier model Gemini 3.5 Flash. Managed Agents eliminate the need for complex orchestration code and environment setup. Developers do not need to write lengthy “glue code”; they only need to write two Markdown-formatted versioned configuration files—AGENTS.md and SKILL.md—in their project to directly define the agent’s system roles and specific skills, and then register it as a custom managed agent. This cloud-based temporary Linux environment gives agents a high degree of physical activity boundaries. The agent can not only perform multi-step reasoning, task planning, and tool invocations within the sandbox, but also independently execute code, read and write files, and call external networks to fetch real-time data. Each interaction automatically saves or restores the current sandbox environment state, making it easy to seamlessly recover the session for subsequent API calls and ensuring that all files and runtime state are intact. This managed approach is very similar to Anthropic’s Claude Managed Agents released in April. The two major model giants are converging on agent infrastructure; both are trying to move from being mere API model providers to becoming container-cloud platforms that handle agent security sandboxes and session state. At present, Managed Agents are available in preview through Gemini API, and developers can quickly get started using the custom templates in the Google AI Studio Playground. Meanwhile, Google has also designed a brand-new Gemini API agent skill specifically to help agents build and manage custom agents on their own. For enterprise users, the Gemini Intelligent Agent Platform (Gemini Enterprise Agent Platform) has also opened a private preview. (Source: BlockBeats)
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • 7
  • 8
  • Share
Comment
Add a comment
Add a comment
GateUser-cf218ace
· 6h ago
The messages from Beating monitoring have always been precise, and the paradigm for developing intelligent agents is really about to undergo a major transformation.
View OriginalReply0
GateUser-8ca669fd
· 6h ago
I/O 2026's release pace is good, I'll waitlist for the Managed Agents preview first.
View OriginalReply0
ShortPositionsAtTheElevator
· 6h ago
Cloud-based transient invocation + isolated Linux, both security and convenience at once, killing competitors.
View OriginalReply0
NeonVortexTunnel
· 6h ago
Configuration files are the infrastructure; this idea should have been popularized long ago. Google is finally keeping up with the times.
View OriginalReply0
GateUser-83a2dd8a
· 6h ago
Say goodbye to server operations management? Sounds great, but are there hidden costs when it comes to implementation?
View OriginalReply0
TheRedTelephoneBoothInTheRuins
· 6h ago
Gemini 3.5 Flash mid-range model runs intelligent agents, and the cost-performance ratio is well balanced.
View OriginalReply0
VolcanicMonolith
· 6h ago
Google has finally woken up, and their move to host intelligent agents was brilliantly played, saving developers countless headaches in operations.
View OriginalReply0
  • Pinned