If you are building autonomous agents on the Gemini API, the July 28 update is a practical one. Philipp Schmid and Mariano Cocirio of Google DeepMind announced four new Managed Agents capabilities: a default-model upgrade, environment hooks, cost and scheduling controls, and a free tier. The positioning builds on the previous release (background tasks and remote MCP integration) — within the Gemini Interactions API, a single call coordinates reasoning, code execution, package installation, file management, and web retrieval inside an isolated cloud sandbox.
Default Model: Gemini 3.6 Flash, Zero-Code Upgrade
The antigravity-preview-05-2026 agent now runs Gemini 3.6 Flash by default — the balanced model for reasoning, coding, and tool use — with no code changes required; your next interaction picks it up automatically. Cost-conscious teams can pass agent_config.model to pin explicitly: Gemini 3.5 Flash (the previous generation for general agentic workflows) or Gemini 3.5 Flash-Lite (the lowest latency and cost in the 3.5 family). Model choice becomes a per-agent deployment parameter rather than a platform-wide lock-in.
Environment Hooks: Block, Lint, and Audit Tool Calls Inside the Sandbox
This is the centerpiece. Environment hooks let you run custom scripts before or after every tool call the agent makes inside its sandbox: drop a .agents/hooks.json into the environment, and the runtime executes your handlers on pre_tool_execution and post_tool_execution events. The matcher field supports regular expressions — target multiple tools with |, or catch everything with *.
A representative configuration: a security-gate group runs gate.py before every code_execution or write_file call — if the script returns {"decision": "deny", "reason": "..."}, the call is skipped and the rejection reason enters the model’s context. An auto-format group runs auto_lint.py after every tool finishes to enforce styling. Hooks also support http-type handlers that POST directly to an external endpoint, which is how you route sandbox events into your own observability or approval stack; failure-handling semantics for HTTP hooks are covered in the hooks documentation.
The official post includes a compelling production case: Offdeal, an AI-native investment bank. Banker-ready decks often contain 30+ company logos across buyer tables, sponsor columns, and tombstone grids — and every logo must be the right company, appropriately sized, name-bearing, on a transparent background, and high-contrast on white slides. Founder & CTO Alston Lin puts the before-and-after plainly: the sandbox is remote, so previously their validation code had nowhere to run. With hooks, a post_tool_execution hook triggers their pipeline inside the sandbox the moment the analyst agent writes its company list — fetching candidates, running pixel-level quality checks, verifying each logo with Gemini vision, and publishing a manifest of approved files as the only images allowed into the deck. The old managed-agents excuse — “custom validation can’t run in a remote sandbox” — is dismantled by this mechanism.
Cost Control: Budget Caps and Resumable Pauses
Managed agents run multi-turn autonomous loops, and complex tasks can burn through token budgets. You can now pass max_total_tokens in agent_config (covering input + output + thinking). When the limit is hit, execution pauses safely and the interaction returns status: "incomplete" — environment state is fully preserved, and passing previous_interaction_id with a fresh budget continues exactly where it stopped. This turns a spent budget from an error into a recoverable pause, which matters in production.
The free tier is open too: Managed Agents now work on free-tier projects without active billing, using just an API key — enough to experiment with agentic workflows.
Scheduled Triggers and Environment Management
Scheduled triggers bind an agent, environment, prompt, and cron schedule into a persistent resource that fires without manual intervention; each run reuses the same sandbox, so files persist across executions. The companion Environments API lets you list, inspect, and delete sandbox sessions from code — recover environment IDs after a disconnect, or clean up sandboxes when your pipeline finishes instead of waiting out the 7-day TTL.
What This Means for Builders
Together, these updates push Managed Agents from “a service that runs your agent” toward “a governable autonomous worker”: hooks close the governance gap (validation code finally has a place to run), budget caps turn runaway risk into a resumable pause, and triggers plus environment management make long-lived scheduled workers viable. Compared with self-hosting your agent infrastructure, Google is selling the managed answer to the same problem set — whether to buy in depends on whether your validation pipelines and data boundaries can live inside someone else’s sandbox. The hooks mechanism meaningfully shrinks that objection — but http handlers still egress to endpoints you control, so the trust question shifts from “can code run there” to “who sees the events.”
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.