Ian Ballantyne and KP Sawhney from Google DeepMind discuss their internal tool Antigravity, an integrated development environment that manages multiple AI agents to automate coding and other tasks while emphasizing human-in-the-loop oversight for quality control. They highlight challenges in scaling agentic systems, optimizing resource use, and developing observability tools, while exploring future directions for agent collaboration, evaluation, and integration within large organizations.
In this panel discussion, Ian Ballantyne and KP Sawhney from Google DeepMind introduce their work on agentic software and demonstrate the capabilities of their internal tool, Antigravity. Antigravity is an integrated development environment (IDE) with a built-in agent manager that allows spawning and managing multiple AI agents working on different coding projects simultaneously. The demo showcased how the tool can analyze specifications, generate code, interact with web pages by inspecting the DOM, and provide detailed reports including screenshots or videos of the tasks completed. This highlights the human-in-the-loop approach where users can review, edit, and guide the agents’ work.
KP Sawhney elaborates on his role in developing the Deep Research Agent and scaling the Antigravity harness to support a wide range of tasks beyond coding. He emphasizes the complexity of managing a large monorepo and the importance of enabling collaboration between different parts of the AI pipeline through shared resources. The focus is on refining the system to be efficient and versatile, supporting various internal use cases at DeepMind and Google, including accelerating code delivery and building a robust library of reusable skills.
A significant challenge discussed is scaling agentic systems to support many users while managing resource consumption, especially given the high token usage of large language models. The team is actively working on optimizing costs and efficiency, including mixing different models like Gemini 4 and more advanced ones to balance performance and resource constraints. They also address the difficulty of evaluating complex agent workflows and the need for custom observability tools to monitor agent behavior, detect issues like looping, and ensure smooth operation at scale.
The conversation touches on the future of agent orchestration, including how multiple agents might collaborate on tasks and how humans can supervise these digital workflows effectively. Both speakers acknowledge the evolving nature of agent architectures and the importance of maintaining control over agent behavior, especially in large organizations. They also discuss the integration of skills and guardrails to enhance agent reliability and the ongoing exploration of better evaluation methods and tooling to support agent development and deployment.
Finally, the panel addresses questions about managing usage limits, pricing models, and code review automation. They highlight the need for seamless fallback mechanisms when users hit model quotas and the potential for agents to assist in code review by adhering to style guides and providing suggestions. The discussion concludes with reflections on the growing volume of AI-generated code and the necessity of building infrastructure to support this shift, emphasizing that while agents can handle repetitive tasks, human oversight remains crucial to ensure quality and effectiveness.