Gemini API managed agents add a free tier and cron triggers
By Steven Van ·
Google adds token budget caps, resumable interactions, and scheduled recurring runs to the managed agents API.
Google shipped the update that turns Gemini API Managed Agents from a demo into infrastructure: free tier availability, budget guardrails, and scheduled triggers. Managed agents already let you call a single endpoint while Gemini handles reasoning, code execution, package installs, file management, and web access inside an isolated cloud sandbox; these additions answer the three questions every developer asked next, what does it cost to try, how do I stop runaway spend, and how do I run it on a schedule.
Free tier: agents without a credit card
Managed agents preview access now extends to free tier projects. Use an API key from a project without billing enabled and interactions run under the free rate limits and usage quota instead of charging you. That's a meaningful on-ramp: autonomous agents are exactly the kind of API you want to prototype before committing a budget to, and until now the sandbox execution model made that awkward.
Budget guardrails: incomplete beats bankrupt
An agent loop of reasoning, tool calls, and code execution can burn tokens fast across multiple turns. The new max_total_tokens field in agent_config caps the total (input, output, and thinking) an interaction can consume. Hit the cap and the interaction stops safely with status: "incomplete", preserving the agent's work and the sandbox filesystem. The elegant part is resumption: pass previous_interaction_id and the environment with a fresh budget, and the agent continues right where it stopped. Runaway tasks become pausable tasks, monitored live over SSE streaming.
Triggers: cron for agents, with memory
Scheduled triggers bind an agent, environment, prompt, and cron schedule into a persistent resource that fires on its own, no external scheduler, no dedicated infrastructure. The key design choice: each scheduled run reuses the same sandbox environment, so files created or repos cloned in one run are there for the next. That makes triggers genuinely useful for recurring work like daily issue triage, nightly regression reports, or repo maintenance, with network allowlists (including header transforms for tokens) controlling what the agent can reach.
Why it matters
Together the three features describe a production posture: cost-controlled, automated workers operating in real dev environments without external orchestration. It's also a direct competitor to running your own agent loop, and, notably, Google published an agent-installable skill for the API itself (npx skills add google-gemini/gemini-skills), the same skills-as-distribution pattern PayPal used on Replit this same day.
Get started
The examples use the @google/genai SDK; the Gemini Interactions API overview and managed agents quickstart at ai.google.dev cover custom agent definitions, environment config, network rules, and streaming patterns. A free-tier key plus a 10,000-token budget is a safe first experiment.
Sources: Google AI Studio on X, Gemini API Docs.