How Google DeepMind Runs Agents at Scale
summary
Demo shows an Antigravity agent manager that spawns multiple agents, edits code, inspects the DOM and returns screenshots or videos; internal policy enforces worse token quotas for employees so customers get priority; teams run a Darwinian skills library where contributions compete and only top skills persist; pipeline work aims at a shared file system replacing context blobs to let pipeline components collaborate; observability uses an agent trajectory store, mock TPUs, and automated code review.
AI-written summary · about
chapters (13)
- 0:00Introduction of KP Sawhney and Ian Ballantyne
- 0:57Demo of the anti-gravity agentic platform
- 4:32Discussion on KP's work with the deep research agent
- 5:46Using skills libraries and managing agent sprawl
- 7:52Addressing scalability and token quotas at Google
- 9:44Audience Q&A: Managing per-user agent behaviors
- 13:12Audience Q&A: Observability and agent trajectory stores
- 14:46Audience Q&A: Future of deep research pipelines
- 16:10Audience Q&A: Handling multi-agent systems
- 17:45Audience Q&A: Perspectives on skills vs. MCP
- 19:12Audience Q&A: Evaluating agentic workflows
- 20:25Audience Q&A: Handling model limits and quota management
- 22:48Audience Q&A: Automated code review processes
related talks
25mDefying Gravity
24mAI Agent Swarms Are Amazing
25mFrontiers of Agentic AI
25mDirecting a Swarm of Agents for Fun and Profit
01h 27mBuilding Vertical Agent with AG2 AgentOS
19mAgents in Action: How Google is Automating Operational Toil
30mMind the Gap - A Developer's Roadmap To Building AI Agents
31mBuilding, Deploying and Managing AI Agents at Scale