Google has added spending caps, code-inspection hooks and scheduled cron triggers to Managed Agents in the Gemini API, the interface that lets a single call spin up a cloud sandbox where an agent can reason, write code, install packages and browse the web. The company detailed the update in a July 28 post from Philipp Schmid and Mariano Cocirio of Google DeepMind. None of these additions make the agent smarter. They make it safer to leave alone.
That distinction matters more than the model swap buried in the same announcement. Managed Agents now default to Gemini 3.6 Flash, replacing the prior generation, and developers can instead pin to Gemini 3.5 Flash or the cheaper Flash-Lite variant. Model selection is a routine update. The other three changes are not.
The more consequential addition is the hooks system. Developers can drop a hooks.json file into the sandbox that runs a custom script every time the agent tries to execute code or write a file, before the action happens. If that script returns a denial, Google says the tool call never executes and the agent is told why it was blocked. A second hook type fires after every action to catch formatting or quality problems the agent’s own output missed.
Google cites Alston Lin, founder of the AI-focused investment bank OffDeal, describing a post-action hook that checks every company logo an agent inserts into a client deck for correct sizing, a transparent background and brand accuracy before the file is allowed into the final document. That is the kind of control an agent needs before anyone lets it touch client-facing work unsupervised: not a smarter model, but a checkpoint that can veto one specific action and say why.
The second addition is a hard ceiling on token spend. Developers can pass a max_total_tokens parameter that caps input, output and reasoning tokens combined for a task. When an agent hits that limit, Google says the run pauses in an incomplete state rather than failing outright, and a developer can resume it later by supplying the paused interaction’s ID along with a fresh budget.
Paired with scheduled triggers, which bind an agent, a prompt and a cron schedule so a task recurs without a person starting it each time, the budget cap turns a one-shot coding assistant into something closer to an unattended worker with a spending limit built in. Google is also opening Managed Agents to free-tier projects, so developers can test agentic workflows before attaching a billing account.
None of this is glamorous. A hook that lints a file or a parameter that halts a run at a fixed token count will not headline a keynote next to a benchmark chart. But they answer the question that decides whether an agent can run without a person watching every step: what happens the moment it tries something nobody anticipated, and who is accountable when it burns through more compute than the task justified. Anyone who has left an autonomous coding agent running unsupervised for an afternoon knows the failure mode is rarely a wrong sentence. It is an unreviewed command, or a bill that clears before anyone checks the dashboard.
For teams building production agents on Gemini, the practical takeaway is to treat hooks.json and max_total_tokens as required configuration rather than optional extras before any agent gets write access to a real repository or a live budget.
Google detailed these Gemini API Managed Agents updates in a July 28, 2026 blog post by Philipp Schmid and Mariano Cocirio of Google DeepMind.
