How Google DeepMind Runs Agents at Scale — KP Sawhney & Ian Ballantyne, Google DeepMind
Summary
KP Sawhney (software engineer on DeepMind's AI platform team) and Ian Ballantyne (developer relations engineer at Google DeepMind) lead a panel at AI Engineer Europe 2026 covering how Google DeepMind builds and scales agentic software internally. The session opens with a live demo of Antigravity, DeepMind's internal agentic IDE platform, which combines a Visual Studio-style interface with a full agent manager framework. The demo shows Antigravity spawning an agent to implement a game spec: the agent analyzes existing files, inspects the DOM, creates an implementation plan, edits it collaboratively with human-in-the-loop feedback, and ultimately rewrites the game from scratch using a Gemini model. The agent captures screenshots and videos of runtime behavior and maintains a scratchpad of reasoning traces visible to the operator.
KP describes his current focus: evolving the Deep Research agent pipeline. The current architecture passes large context blobs through each stage of the research pipeline, which is expensive and consumes context window capacity. His proposed replacement replaces these blobs with a shared file system so each pipeline component can collaborate like human researchers working in a shared workspace, and can produce supporting artifacts such as infographics and structured documents that the current architecture cannot generate.
On scaling agents inside Google, the speakers highlight token quota management as the primary operational challenge. DeepMind's internal users have worse quota limits than paying Gemini customers — by design, to protect customer SLAs. Power users who spike cluster usage are identified by SRE teams monitoring graphs 24/7, who call teams directly and ask them to stop. KP acknowledges this is currently brute-force quota enforcement and notes that future pricing models need to evolve: the flat subscription model does not work well for token-hungry agentic workloads, citing Anthropic's move to block automated Claude usage as evidence that the industry is grappling with the same problem.
The skills library is described as a Darwinian system: thousands of engineers contribute skills to a central library, but only the strongest survive through a quality filtering process. Contributors with deep domain expertise author skills that other engineers and their agents then consume for free. KP notes he has skills for debugging raw logs, code review style checking, and other engineering tasks. He expresses skepticism about MCP as a long-term standard, preferring the combination of skills plus guardrail CLI interactions, though he acknowledges DeepMind supports both.
On observability, DeepMind uses a custom web app that captures every agent query issued on its central agent backend, allowing operators to drill down through system hierarchy to raw model predict requests. A separate agent trajectory store is used specifically for coding agents, enabling engineers to diagnose exactly where looping or off-rails behavior began. On evaluation, KP identifies the hardest problem as generating new domain-specific datasets and setting up the sandboxed environments needed to run them — noting that agents themselves are being experimented with to author test cases for their own skills.
For automated code review, DeepMind uses per-language auto-review models fine-tuned on internal style guides and historical good-code examples, supplemented by product-area-specific SRIs and prompts. KP recounts receiving an unsolicited agent PR comment with a high-quality suggestion — triggered by a colleague's autonomous agent, not by KP himself. The panel closes with discussion of model fallback under quota exhaustion: the vision is for the harness to seamlessly downgrade from a pro model to Flash, then to a local model, without interrupting the workflow or returning a silent failure after an hour of inactivity.
Key Takeaways
- •Replace large context blobs in research pipelines with a shared file system so each pipeline component collaborates like a human researcher, enabling artifact generation (infographics, structured documents) that blob-passing cannot support.
- •Internal users at Google deliberately receive worse token quotas than paying customers — enforce this by design and have SRE teams proactively reach out to teams spiking cluster usage rather than relying on automated hard-stops alone.
- •Apply a Darwinian filter to skills libraries in large organizations: allow mass contribution but ensure only high-quality skills survive, so the library remains useful rather than sprawling into unmaintained noise.
- •Build a custom agent trajectory store for coding agents that captures each reasoning step, enabling precise diagnosis of where looping or off-rails behavior started — generic logging is insufficient for this.
- •Mix cheap open-weight models (e.g., Gemma 4, effectively free on owned GPUs/TPUs) with advanced models only for specific high-value pipeline steps to dramatically reduce agentic workflow cost.
- •Design quota exhaustion to trigger seamless model fallback (pro → flash → local model) rather than silently terminating a long-running job and returning an error after an hour of inactivity.
- •Use mock TPUs or sandboxed environments to evaluate agentic harness flows without consuming real TPU budget — separate harness testing from model quality testing.
- •Per-language auto-review models fine-tuned on internal style guides and good-code examples provide consistent, scalable code review signal at the pace of agent-generated PRs.
- •MCP's primary value is in the auth layer; for many internal agentic workflows, a combination of skills plus guardrail CLI interactions may outperform MCP-based tool integrations.
Transcript
Okay. Hello everybody. So, my name is N Ballantyne. I'm a developer relations engineer at Google DeepMind. Hi folks. I'm KP Sawney, software engineer in DeepMind's AI platform team. We're going to do an agentic panel today to talk a little bit about how DeepMind thinks about agentic software, how we build kind of our own stacks. We're going to start very briefly by just showing a quick demo and then we'll go into like a like a discussion about some of the stuff that KP works on and hopefully we ...
Related Sessions
