Prototyping as Leadership: Scaling Output with AI Agents

The Overnight Execution Loop

Hursh Agrawal argues that the manager schedule is now viable for active development by leveraging autonomous coding agents to handle execution during off-hours. The core workflow relies on a 5:00 PM setup block where the leader provides the agent with deep business context—often gathered by a secondary agent that scrapes Slack, Notion, and Jira—to define the task scope. The agent then operates autonomously overnight, writing tests first, implementing features, and running internal AI code reviews to ensure the resulting pull requests are ready for morning inspection.

Hill Climbing and Model Training

Beyond feature development, this loop is used to optimize AI performance and train custom models. For feature optimization, the author collects feedback dumps (JSON files) from users, which are then used as an eval set for an overnight hill-climbing process. The agent iteratively refines system prompts or logic until the performance metrics improve. Similarly, the author trains custom classifiers (e.g., PII detection) by providing training data and AWS access to an agent, which then selects the model architecture, trains the model on an EC2 instance, and prepares the inference code for production.

Scaffolding and Leadership Hygiene

This approach requires significant organizational infrastructure to remain safe. The author emphasizes the necessity of robust CI pipelines, sophisticated feature flagging, and a dedicated prototype branch to prevent production incidents. Leaders must maintain high standards for PR hygiene, as their output models behavior for the rest of the engineering team. Crucially, the author warns against adding reviewers to code that has not been personally read, noting that agents are not perfect and can produce obvious errors that damage a leader's credibility if pushed to the team without verification.