Google opened the agent harness that powers Antigravity to anyone with a Gemini API key. Managed Agents in the Gemini API let a single call provision a fully-formed agent inside an isolated, ephemeral Linux sandbox—the same environment Google runs internally for its own coding agent and Deep Research—configured through versionable markdown files instead of orchestration code.
One Call, One Sandbox
The shape of the API is the headline. You hit the Interactions API with agent: "antigravity-preview-05-2026" and a user input, and Google provisions a remote Linux environment with file system, code execution, browser, and tool calling already wired up. The response returns an Interaction object with an id and an environment_id; passing those back into the next call keeps the agent operating in the same sandbox across turns.
The Antigravity agent itself runs on the new Gemini 3.5 Flash, which Google says outperforms Gemini 3.1 Pro on most benchmarks while running roughly 4× faster—the speed needed for an agent loop where every turn pays a planning, tool-use, and synthesis cost. Each interaction can chew through 100K to 3M tokens; environment compute is not billed during the public preview.
AGENTS.md and SKILL.md as Source Files
The configuration story is where this gets interesting for developers. Rather than wiring custom agents through bespoke orchestration code, Google leans on a pattern that's becoming a de facto standard: agents and skills defined as markdown files you check into a repo. AGENTS.md describes the agent's instructions, defaults, and posture. SKILL.md files describe individual skills the agent can call. Both are registered as part of a managed agent and versioned like any other source file.
The same naming convention is what Next.js ships in create-next-app and what Anthropic uses for Skills. Google is reading from the same playbook, which means the markdown configs you write for one ecosystem now travel further than they used to.
Context Compaction Built In
One of the chronic problems with long-running agents is context rot—reasoning steps, tool traces, and file contents piling up until the model loses focus and the run derails. Google's Managed Agents ship with native context compaction at roughly 135K tokens, automatically summarizing and pruning history so the agent can keep going without hitting limit errors or degrading.
That's a feature that has been left to library authors and harness builders for the last 18 months. Putting it inside the API means every Gemini agent gets the same compaction logic Google uses on Antigravity, with no LangChain or LangGraph plumbing required—though those frameworks remain supported alongside CrewAI, LlamaIndex, the Vercel AI SDK, Google ADK, and a new Python Antigravity SDK.
How It Stacks Up
Managed Agents put Google on parity with Anthropic's Claude Managed Agents and OpenAI's workspace agents—the third major lab to ship a hosted, sandboxed agent runtime in eight weeks. The three offerings are converging on the same shape: provision a sandboxed environment, give the agent tools and skills, run multi-turn interactions with built-in memory management, and bill for tokens plus tool usage.
Google's twist is the depth of the existing infrastructure underneath. Antigravity, Deep Research, and the new Antigravity 2.0 desktop all run on this harness, which means the API isn't a new product—it's a productized version of the system Google has already been running at scale. Enterprise customers also get Managed Agents on the Gemini Enterprise Agent Platform in private preview.
Why It Matters for Web Developers
The infrastructure tax for building a real agent has been brutal: sandbox provisioning, tool routing, context window management, retry logic, observability, the whole stack. Managed Agents in the Gemini API collapse that into one API call and two markdown files. If you've been waiting to ship a real agent feature because the plumbing wasn't worth it, the plumbing just got handled.
The bigger pattern across this week's announcements—Composer 2.5, workspace agents, Claude Managed Agents, and now Gemini Managed Agents—is that the agent harness is no longer the differentiator. The differentiator is what you point the harness at: which skills, which tools, which data, which workflows. The labs are racing to make the harness boring, and that's good news for everyone building on top.