Google Gemini API adds managed agents, sandboxes

// 9h agoPRODUCT UPDATE

Google Gemini API adds managed agents, sandboxes

Google's new Managed Agents and Interactions API provide secure, stateful Linux sandboxes for LLMs to execute code and browse the web autonomously. Developers can now deploy versioned agents defined by markdown files without managing underlying execution infrastructure.

// ANALYSIS

Google is moving Gemini from a simple chat API to a full-stack agent orchestration platform by internalizing the most difficult part of agentic workflows: secure code execution.

–Managed sandboxes eliminate the infrastructure overhead and security risks of running LLM-generated code in production.
–The Interactions API replaces the stateless `generateContent` model with server-side session management, reducing token overhead for long-running tasks.
–Using `AGENTS.md` and `SKILL.md` for configuration treats agent identity and tool-use as version-controlled code rather than opaque prompt strings.
–Native web browsing and code execution capabilities position Gemini 3.5 Flash as a direct competitor to specialized "coding agent" platforms.
–Server-side state persistence enables "Deep Research" style background tasks that survive client disconnects.

// TAGS

gemini-apiagentsandboxingcode-executiondevtoolapimcptool-use

DISCOVERED

9h ago

2026-05-19

PUBLISHED

9h ago

2026-05-19

RELEVANCE

9/ 10

AUTHOR

Wes Roth

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE4h ago

FiveThirtyEight Index hits Internet Archive

Journalist Ben Welsh launched a searchable index of over 22,000 FiveThirtyEight articles preserved on the Internet Archive's Wayback Machine. The project restores access to 15 years of data-driven reporting after the original site archive was taken offline by corporate owners.

OPEN SOURCE5h ago

Karpathy-inspired CLAUDE.md hits 137K stars

multica-ai/andrej-karpathy-skills is a viral configuration framework for AI coding agents that enforces strict behavioral guardrails based on Andrej Karpathy's observations of LLM pitfalls. The single CLAUDE.md file transforms vague imperative instructions into surgical, verifiable goals to eliminate common agentic hallucinations.

MODEL6h ago

Google launches Gemini Omni world model

Gemini Omni is a native multimodal foundation model that enables conversational video editing through natural language. It understands real-world physics and motion to modify scenes, characters, and lighting while maintaining perfect temporal continuity.