How Google Built Agent Smith: An Asynchronous Coding Agent That Works While Engineers Sleep
Google has been quietly running one of the most instructive large-scale experiments in agentic coding, and Business Insider obtained the first detailed inside account. "Agent Smith" — built on top of Google's internal Antigravity agentic platform — has become so popular within the engineering organization that access had to be restricted to manage server load. That detail alone is worth sitting with: this isn't a demo tool, it's one engineers are actively fighting to use more of.
What distinguishes Agent Smith from earlier coding assistants is the execution model. Rather than generating code in response to a prompt and waiting for the next input, Agent Smith plans and executes multi-step workflows autonomously, with access to internal tools and the ability to pull documents from an employee's profile that they would otherwise need to retrieve manually. The critical architectural detail is that it runs asynchronously: engineers hand off a task from their phones, disconnect, and the agent works in the background without requiring an active laptop session. Engineers then check in on progress and redirect as needed — a fundamentally different workflow from babysitting each generation.
The broader organizational context matters too. Sergey Brin named agents as a major 2026 strategic focus at a recent company town hall, and Google leadership has shifted from encouraging AI tool use to treating it as a baseline expectation for coding roles. The Antigravity-to-Agent-Smith architecture — where a general agentic platform acquires specialized autonomous capabilities — mirrors what many enterprise engineering teams are building right now. The async delegation workflow it represents isn't Google-specific; it's a pattern available to any team with a capable coding agent today.