Google Just Shipped a Managed Agents API That Spins Up a Full Linux Sandbox With One Call. The Infrastructure Moat for Building Agents Is Gone.
At Google I/O 2026 this week, Google announced Managed Agents in the Gemini API. The pitch is simple: one API call, and you get a fully provisioned agent that reasons, uses tools, executes code in an isolated Linux environment, and cleans up after itself. No infrastructure to configure. No sandbox to manage. Just an API call and a running agent.
Gemini 3.5 Flash powers it. Google claims it outperforms Gemini 3.1 Pro on nearly all benchmarks while running four times faster than competing frontier models. Whether that holds up in production is a different question, but the infrastructure play here isn't really about which model is faster.
The infrastructure play is about who gets to build agents.
What managed agents in the Gemini API actually do
The Gemini Managed Agents API (built on what Google internally calls the Antigravity agent harness) handles the parts of agent infrastructure that eat weeks of engineering time: isolated execution environments, tool call routing, error recovery, context management across long-running tasks.
Before something like this existed, building an agent product meant standing up sandboxes (E2B, Modal, your own Docker fleet), writing retry logic for tool failures, deciding how to handle agents that hit rate limits mid-task, and managing the accumulated context of multi-step runs. That's 2–4 weeks of infrastructure work before you've written a single line of product logic.
With a single API call, Managed Agents provisions all of that. The Linux sandbox is remote, isolated, and ephemeral. You pass it a task, give it tools, and it runs. You get the result.
This is the same architecture Anthropic shipped with Claude Managed Agents: Anthropic announced their self-hosted sandbox variant just three days before Google I/O. Both companies are converging on the same insight: the infrastructure barrier is what's keeping agents from being as easy to build as chatbots were in 2023.
Antigravity 2.0 and what it does beyond the API
Google also shipped Antigravity 2.0, a desktop application and CLI for developers who want to run agents locally or across complex multi-agent workflows. The new version adds specialized subagents for different workflow components, built-in cross-platform terminal sandboxing, credential masking, and hardened Git policies.
I'd frame Antigravity 2.0 as the Claude Code equivalent for the Google ecosystem. If you're already in the Google stack (using Firebase, Cloud Run, BigQuery), Antigravity is probably the fastest path to agent-assisted development. For developers who aren't committed to either platform, you now have two real options with meaningfully different pricing structures and API designs.
Google is also launching a $100/month AI Ultra plan with 5x higher usage limits than the $20 AI Pro tier, explicitly targeting developers and technical leads. That's the pricing tier where you'd run Managed Agents at any real volume.
The signal this sends about where agents are going
Six months ago, building a production-ready agent workflow was a genuine engineering challenge: not because the models weren't capable, but because the infrastructure to run them reliably was either expensive, fragile, or both. Teams that shipped good agent products had a real advantage.
That advantage is compressing. Fast. When Google and Anthropic both ship one-call managed agent infrastructure in the same week, the differentiation stops being "can you build an agent at all" and becomes "what are you doing with it." The commodity layer is the sandbox. The value is in the product logic on top.
For a solo operator, this is unambiguously good. You can now build what required a 3-person team six months ago. The question is whether you use the freed-up engineering time to build something interesting, or whether you just build the same thing everyone else is building with the same tooling.
The honest counter-take
Managed agent APIs at Google-scale introduce a new risk: vendor-managed infrastructure means Google controls the execution environment your product runs in. When something breaks, and it will break, your debugging surface is the API's error response, not a stack trace from a sandbox you own.
This is the same trade-off you make with any managed service. It's usually the right trade-off. But it's worth being clear that "one API call" is a convenience layer on top of real infrastructure, and the failure modes of that infrastructure are now someone else's problem, and therefore, partially out of your control.
The $100/month AI Ultra tier is also a meaningful price step if you're running agents at volume. Managed agent execution isn't free: compute in sandboxes costs real money, and at some usage threshold, running your own sandbox infrastructure on Modal or Fly.io will be cheaper. Know your usage pattern before committing.
What I'd actually do
If you've been putting off building an agent product because the infrastructure setup felt like too much overhead, this week's announcements from both Google and Anthropic are the clearest sign yet that the setup cost is going to zero. Start building.
Pick the platform where you're already spending money. If you're using Claude for everything else, Anthropic's Managed Agents is the path of least resistance. If you're already in GCP, Gemini's API will integrate more cleanly with your existing auth and billing. Don't pick based on benchmark scores, pick based on where the integration friction is lowest for your specific stack.
The models are commoditizing. The infrastructure is commoditizing. The product is the thing.
Author
Lukas
@lukcombinator