Google Just Made Production-Grade AI Agents Accessible Through the Gemini API

---

Announced at Google I/O 2026, Managed Agents in the Gemini API is one of the more developer-consequential releases to come out of the event. Google is launching Managed Agents in the Gemini API, enabling developers to spin up an agent that reasons, uses tools, and executes code in an isolated, ephemeral Linux environment with a single call.

The underlying problem this solves is real. Building a production-grade agent previously meant managing complex infrastructure, scaffolding, and isolated sandboxes to serve customers at scale. With Gemini Managed Agents, Google is abstracting away that complexity so developers can focus on product experience and agent behavior.

As someone who covers the agent space daily, I can say this is one of the cleaner infrastructure abstractions I've seen from a major AI provider. The configuration-first approach, in particular, deserves a close look.

---

What the Antigravity Agent Actually Does

The Antigravity agent is a general-purpose managed agent on the Gemini API. A single API call gives you an agent that reasons, executes code, manages files, and browses the web inside your own secure Linux sandbox, hosted by Google.

The Antigravity agent (antigravity-preview-05-2026) is the general-purpose managed agent. It can execute code in Bash, Python, and Node.js, manage files, browse the web, and use Google Search.

Here's what a minimal invocation looks like via the Interactions API:

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "Content-Type: application/json" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Api-Revision: 2026-05-20" \
  -d '{
    "agent": "antigravity-preview-05-2026",
    "input": [{"type": "text", "text": "Write a Python script that generates the first 20 Fibonacci numbers."}],
    "environment": "remote"
  }'

Unlike a standard chat request that produces a single output, an Antigravity interaction is an agentic workflow. A single request triggers an autonomous loop of reasoning, tool execution, code running, and file management.

---

AGENTS.md and SKILL.md: Configuration as Code

The most interesting design choice here is how Google handles agent customization. Instead of requiring developers to write orchestration code, they opted for a filesystem-native, markdown-driven approach.

You can extend the Antigravity agent with your own instructions and skills. Instead of writing complex orchestration code, you can define everything in markdown files like AGENTS.md and SKILL.md and register them as a managed agent.

The agent supports a filesystem-native approach to customization: you can mount files like AGENTS.md for instructions and skills under .agents/skills/ directly into the sandbox, or pass configuration inline at interaction time. You can iterate on your configuration inline and then save it as a managed agent when you are ready.

This matters for teams that care about version control. Agent definitions become plain text files you can diff, review, and roll back like any other source artifact. Developers can define custom agents through versionable markdown files such as AGENTS.md and SKILL.md, rather than building complex orchestration layers from scratch. Each managed agent gets its own ephemeral sandbox provisioned with skills, Model Context Protocol (MCP) servers, and server-side tools.

---

Key Technical Highlights

Three capabilities define this feature: first, the agent harness itself — the same technology and infrastructure that powers Google's own agents, co-optimized with Gemini 3.5 Flash. Second, persistent isolated environments — each interaction creates an environment that can be resumed in follow-up calls with all files and state intact, enabling seamless multi-turn sessions without reinitializing context. Third, custom agent definitions — developers can extend the Antigravity agent with custom instructions and skills using markdown files, with new custom agent templates available in the Google AI Studio Playground.

A few additional constraints worth knowing before you build:

• The Antigravity agent does not support structured outputs. Tools like file_search, computer_use, google_maps, function_calling, and mcp are not yet supported.
• Audio, video, and document inputs are not supported at the moment. Only text and image are allowed.
• Unlike standard Gemini models, the Antigravity agent runs through multiple autonomous loops per interaction and can accumulate a high number of tokens. You can monitor agent runs through SSE streaming and cancel the request if the agent appears stuck.

---

Pricing and Token Considerations

Pricing follows a pay-as-you-go model based on the underlying Gemini model tokens and the tools the agent uses.

Costs vary based on task complexity. The agent autonomously determines how many tool calls, code executions, and file operations are needed. That autonomy is a double-edged sword: you get more capable task completion, but you also need to monitor spend carefully on long-running tasks.

50–70% of input tokens are typically cached, which meaningfully reduces costs on multi-turn sessions where the agent re-reads the same context repeatedly.

---

The Model Under the Hood

Gemini 3.5 Flash outperforms Gemini 3.1 Pro across almost all benchmarks while running four times faster than other frontier models, providing the high-speed engine needed for real-world agentic workflows.

According to Google, 3.5 Flash outperforms Gemini 3.1 Pro across almost all benchmarks while running four times faster than other frontier models. The speed advantage is practically significant when multiple agents run in parallel, since model latency compounds across concurrent agent calls.

For agentic workloads, where a single user request might trigger dozens of model calls, that latency reduction isn't cosmetic. It directly affects how fast your product responds.

---

Enterprise Path and What Comes Next

For enterprises, Google has also added support for managed agents in the Gemini API on the Gemini Enterprise Agent Platform in private preview.

What makes this deployable is the Google Cloud sandbox, which is secure by design. The agent harness runs on Google's servers, and each agent has its own ephemeral sandbox provisioned with your skills, MCP servers, and server-side tools.

Google said full integration with A2A and Agent Platform governance and security is coming soon. The Agent Registry, Agent Gateway, and IAM governance policies are all in various stages of preview, filling out the security layer that enterprise teams will need before going to production.

Google said the four approaches — Agent Studio, Managed Agents API, Antigravity, and ADK 2.0 — are additive, meaning agents built at one level can be used as part of more complex systems at another level.

---

Final Thoughts

The AGENTS.md and SKILL.md pattern is the part of this release I find most worth watching. Treating agent configuration as versionable, reviewable source files is a sane engineering decision, and it aligns with how software teams already work. If this pattern gains traction, it could become a de facto convention the way Dockerfiles or GitHub Actions workflows did.

The current limitations are real though. No MCP support, no structured output, and text-plus-image-only inputs means this isn't a drop-in for every use case. The token accumulation behavior in autonomous loops also requires careful cost monitoring in production, especially for open-ended tasks.

Managed Agents in the Gemini API is rolling out in preview, and you can get started instantly with new custom templates in the Google AI Studio Playground. If you're already building on the Gemini API, it's worth spinning up a test interaction this week. What do you think? Drop your thoughts in the comments.

---