How Google DeepMind Runs Agents at Scale — KP Sawhney & Ian Ballantyne, Google DeepMind

IDE-integrated agent manager with browser control

Antigravity provides a VS Code-style interface where agents can spawn multiple instances, control browsers, inspect DOM elements, and capture screenshots or videos of interactions.

Human-in-the-loop workflow

The system generates editable implementation plans and scratchpad notes that allow users to review, modify, or reject agent actions before proceeding.

Built-in planning and tool use

Agents autonomously create to-do lists and utilize internal tools to analyze specifications, write code, and verify implementations through browser automation.

Token hunger as primary bottleneck

Managing per-user and per-team quotas is critical at Google scale to prevent power users from overwhelming systems with token-hungry agentic workflows.

Hybrid model routing strategy

DeepMind mixes free local models like Gemma 4 for general tasks with advanced APIs for specific components to optimize costs and quota efficiency.

Incompatibility of subscription models

Traditional subscription pricing breaks down for agentic systems, necessitating new metering approaches as brute-force quota limits are currently the primary defense.

Shift from monolithic to collaborative architecture

Deep Research is evolving from passing massive text blobs between stages to agents collaborating via shared filesystem workspaces.

Darwinian skill library approach

Internal users contribute to a sprawling library of skills where only the best survive through organizational natural selection, democratizing expert knowledge.

Human-supervised digital assembly lines

The future vision involves humans overseeing parallel agent tracks that communicate efficiently, moving beyond current single-agent limitations.

Custom trajectory stores and tracing

DeepMind built internal tools including agent trajectory stores to diagnose exact failure points, looping behaviors, and hierarchical drill-downs to raw model requests.

Expensive evaluation infrastructure

Testing requires sandboxed environments and mock TPUs to validate agent harnesses without consuming scarce production compute resources.

Skills preferred over MCP

While both are supported, KP Sawhney advocates for skills combined with guardrail CLI interactions over MCP, viewing the latter as potentially transient except for authentication.