Google’s Advent of Agents returns for a third season, this time on security

Google Cloud’s Advent of Agents is running its third season this October, a daily tutorial series built around Gemini, the Agent Development Kit, and the rest of Google’s cloud agent stack. The first two seasons, a 25-day run in December 2025 and a 31-day run in March 2026, left behind an archive of 56 lessons covering the full arc of building an agent rather than just a single demo: giving agents long-term memory, connecting them to outside tools through MCP servers, orchestrating multi-agent setups where one model writes and another critiques the output, adding human approval steps for anything sensitive, and finally deploying the whole thing to Google Cloud infrastructure.

This season narrows its focus considerably. Where the first two runs built toward getting an agent working at all, season three is built around what happens once it’s already in production and something tries to abuse it. The stated theme is governance and security: threat modeling for agent systems, progressive trust models that scale an agent’s permissions based on how much it’s proven itself, and guardrails meant to stop unauthorized data disclosure before it happens. One line from Google’s own framing of the season, that prompt engineering isn’t a security strategy, is a pointed jab at how a lot of teams have been treating agent safety as an afterthought handled entirely through careful wording rather than actual access controls.

The series runs the full 31 days of October, one lesson per day, and it’s free to follow on Google’s own site alongside the archived content from the first two seasons. For teams that have already shipped an agent and are now dealing with the harder problem of keeping it from doing something it shouldn’t, that’s a more useful starting point than another walkthrough of how to wire up a basic chatbot.