SDSignal Desk

Gemini API Managed Agents: 3.6 Flash, hooks, and more

Jul 28, 2026, 9:00 AM · Google AI Blog

Image: Google AI Blog

Managed Agents are becoming a sandboxed worker with policy hooks and budgets, not just a one-shot coding call.

Why it matters

Google says Managed Agents now default to Gemini 3.6 Flash, can pin 3.5 Flash or Flash-Lite, and support environment hooks that run before or after tool calls. A deny decision can skip a call and return a reason into model context.

The update adds free-tier access, a max_total_tokens cap that pauses with incomplete status, cron triggers that reuse a sandbox, and an Environments API for listing and deleting sessions. OffDeal is cited as using post-tool hooks for logo checks.

From the desk

This is Google’s own developer blog, so product claims and the customer example are vendor-reported. The hooks and budget controls are the material change because they acknowledge that autonomous loops can spend tokens and write files.

Signal Desk's read: the sandbox is becoming a control plane. Teams that need linting, audit or spend limits will care more about hooks than about the default model bump.

The remaining gap is operational: pausing on a token cap preserves state, but resuming with a fresh budget still requires the caller to decide whether the agent should continue, retry or stop.

Context

Managed Agents run reasoning, code, packages, files and web retrieval in an isolated cloud sandbox via the Interactions API. Hooks can be command or HTTP handlers.

Who feels it

Developers
Hooks and token caps make longer agent runs more governable.
Finance
Free-tier access and budgets reduce runaway cost risk.
Security
Pre-tool deny paths can block writes inside the sandbox.

What to watch

  1. Adoption of deny hooks in production
  2. How incomplete runs are resumed
  3. Trigger reliability across reused sandboxes

Read the original

Continue at the source.

Google AI Blog

Companies: Google