← back
How Google DeepMind Runs Agents at Scale — KP Sawhney & Ian Ballantyne, Google DeepMind
Takeaway
Scaling agents requires a reusable harness plus disciplined management of skills, context, and inference resources.
Summary
- An Antigravity demo shows multi-agent project management, editable implementation plans, browser control, DOM inspection, and review artifacts.
- DeepMind is exploring reuse of the coding harness for tasks such as research, with shared files replacing large text blobs passed between pipeline stages.
- Internal skill libraries improve productivity but need curation to control duplication and quality drift.
- Scaling agent usage requires per-user and per-team quota management, cost efficiency, and combining smaller models with advanced models for selected tasks.
antigravityagent-harnessdeepmind
Original description
Google DeepMind employees have worse token quotas than paying customers. That is not a mistake. KP Sawhney explains: customers get priority, and if an internal team spikes usage on a cluster someone monitoring 24/7 will just call and ask them to stop. This panel covers how DeepMind thinks about agents at scale from the inside: managing quota across thousands of power users, a Darwinian skills library where only the strongest skills survive as engineers contribute en masse, and where the Deep Research pipeline is going next. KP's current focus is replacing the pipeline's giant context blobs with a shared file system so each research component can collaborate the way human researchers would, and produce supporting artifacts that the current architecture cannot. Speaker info: / ianballantyne https://github.com/irbg / kyle-sawhney https://github.com/KPSawhney Timestamps: 0:00 Introduction of KP Sawhney and Ian Ballantyne 0:57 Demo of the anti-gravity agentic platform 4:32 Discussion on KP's work with the deep research agent 5:46 Using skills libraries and managing agent sprawl 7:52 Addressing scalability and token quotas at Google 9:44 Audience Q&A: Managing per-user agent behaviors 13:12 Audience Q&A: Observability and agent trajectory stores 14:46 Audience Q&A: Future of deep research pipelines 16:10 Audience Q&A: Handling multi-agent systems 17:45 Audience Q&A: Perspectives on skills vs. MCP 19:12 Audience Q&A: Evaluating agentic workflows 20:25 Audience Q&A: Handling model limits and quota management 22:48 Audience Q&A: Automated code review processes