The Case for Leader-Led Prototyping

As AI models evolve every three months, leadership intuition based on second-hand information (like social media) is no longer sufficient. Hursh Agrawal argues that building is now a core leadership function. By dedicating 1-2 hours daily to hands-on development, leaders can calibrate their strategy, understand the actual capabilities of frontier models, and demonstrate new possibilities to their teams through working prototypes rather than abstract arguments.

The Overnight Execution Loop

Agrawal utilizes an "overnight loop" to maximize productivity despite a meeting-heavy schedule. The workflow involves three distinct phases:

  • Context Assembly: Before 5:00 PM, a "coworker agent" is tasked with scanning internal documentation (Slack, Jira, Notion) to generate a comprehensive prompt. This ensures the coding agent has the same business context as the leader.
  • Autonomous Execution: The coding agent (e.g., Claude Code) runs overnight, tasked with writing tests first, implementing features, and managing CI/CD pipelines.
  • Morning Review: The morning block is reserved for reviewing the agent's output, verifying the code, and ensuring the PR meets organizational standards.

This loop extends beyond feature development to include hill-climbing prompts against feedback-driven eval sets and training custom ML classifiers by providing the agent with training data and AWS access to provision sandboxes.

Scaffolding and Hygiene

Building as a leader requires rigorous engineering discipline to avoid becoming a source of technical debt or production incidents. Key requirements include:

  • Organizational Scaffolding: Success depends on trustworthy CI, sophisticated feature flags, and a dedicated prototype branch that allows for internal dogfooding without risking production stability.
  • Code Hygiene: Leaders must model good behavior. This means submitting small, readable PRs and—crucially—never adding reviewers to code that you have not personally read.
  • Verification: Even with autonomous agents, the leader must perform manual testing. AI agents are excellent at execution but still require human judgment to navigate edge cases and business-specific constraints.