Europe 2026
How Google DeepMind Runs Agents at Scale — KP Sawhney & Ian Ballantyne, Google DeepMind
Overview
This talk discusses Google DeepMind's approach to building and scaling agentic software. It highlights the Antigravity framework as an integrated IDE and agent management platform, showcasing its capabilities in code generation, task execution, and human-in-the-loop feedback. The discussion also touches upon the challenges and strategies for managing agentic systems at scale, including resource allocation, cost efficiency, and robust evaluation methods.
Who should watch
- AI engineers and developers building or using agentic systems.
- Product Managers and technical leaders interested in agent capabilities for product development.
- Builders exploring tools and frameworks for large-scale AI agent deployment.
- Those facing challenges with agent scalability, cost management, and evaluation.
Key takeaways
- Google DeepMind utilizes the Antigravity framework, an IDE-integrated platform, for managing and running multiple agents on various projects.
- Agents within Antigravity can analyze specifications, generate code, interact with external tools like browsers, and provide detailed reports on task completion.
- A significant focus is placed on building and refining a library of agent skills, ensuring only the most effective ones are maintained.
- Scaling agentic systems presents challenges related to token consumption and computational cost, necessitating efficient resource management and cost mitigation strategies.
- Strategies for scaling include mixing and matching different model types (e.g., free local models with advanced proprietary ones) and developing efficient evaluation methods, potentially using mock hardware.
- Observability is crucial, with custom internal tools developed for monitoring agent queries, tracing execution, and diagnosing issues like infinite loops.
- Future directions involve making agentic systems more collaborative, akin to human research teams, to improve efficiency and reduce context-heavy data transfer.
- Evaluation of agent performance is complex, requiring robust testing methodologies and the development of new datasets, especially for specific agent skills.
Notable quotes
*The tool itself has like built-in to-dos. It has a like a planning system.*
*My my focus has turned now to making best use of of this anti-gravity harness internally.*
*A big area of focus for us right now is improving those skills, making sure that only the best ones survive really, almost Darwinian nature.*
Unofficial community note. Prefer the recording for nuance.