Detailed Analysis
Boris Cherny, the creator of Claude Code — Anthropic's terminal-based agentic coding tool — has revealed that his personal workflow deploys thousands of AI sub-agents running in parallel overnight to perform what he describes as "deeper work." The disclosure, reported by Business Insider, offers a rare glimpse into how the engineers building frontier AI tools are themselves using those tools at the extreme edge of their capabilities. Rather than relying on single-session, interactive AI assistance, Cherny's approach delegates large volumes of complex, time-intensive tasks to fleets of autonomous agents that operate asynchronously while he sleeps, returning results by morning.
The significance of this revelation lies in what it signals about the maturity and scalability of agentic AI systems. Claude Code, which allows users to give Claude broad access to a codebase and have it autonomously write, debug, test, and refactor code, was itself designed with multi-agent orchestration in mind. Anthropic has publicly discussed the notion of "agent swarms" and hierarchical agent architectures in its technical documentation, but Cherny's disclosure operationalizes that vision at a striking scale. Running thousands of sub-agents simultaneously implies a sophisticated orchestration layer managing task decomposition, parallelization, resource allocation, and result synthesis — infrastructure that goes well beyond typical developer use cases.
This development connects directly to a broader industry shift away from AI as a reactive, prompt-response tool and toward AI as an autonomous, proactive workforce. Major AI labs including OpenAI, Google DeepMind, and Anthropic have all made multi-agent capability a central pillar of their 2025 and 2026 product roadmaps. The notion of "overnight compute" — offloading deep cognitive work to AI agents that operate during off-hours — mirrors how data engineering teams have long scheduled batch processing jobs, but now applies that paradigm to reasoning-intensive software development tasks.
The implications for software engineering productivity are substantial. If the creator of a leading agentic coding tool is himself using thousands of agents to do work he cannot complete in a single workday, it suggests the upper bound of AI-assisted productivity is far higher than most enterprise adopters have yet explored. It also raises practical questions about cost, oversight, and quality assurance: at thousands of simultaneous agents, human review of outputs becomes a bottleneck in itself, pushing toward meta-level AI review pipelines. Cherny's setup likely represents a preview of normalized engineering workflows within AI labs, where the humans directing work and the agents executing it exist in a fundamentally asymmetric relationship of scale.
Read original article →