Sign InOpen Brain
AI EngineerVideoSource Linked

Prototyping as Leadership: How a CTO Ships with AI Agents — Hursh Agrawal, The Browser Company

A manager can turn fragmented time into agent steering: prepare context, run bounded work overnight, then personally test and review small PRs before involving the team.

AI Engineer · Aug 20, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Despite **15+ recurring meetings** and **seven direct reports**, Agrawal reports shipping **2–10 PRs a week** by combining short steering blocks with one overnight agent run. He reserves the work for prototypes, internal tools, codebase gardening, and product vision.

Practical Implication

Gather business and technical context before the run, ask for tests first, exercise the flow with computer use, and have a clean subagent review the result. In the morning, test it yourself, read every change, and split the work into small, reviewer-friendly PRs.

Agent-Ready Context
Despite **15+ recurring meetings** and **seven direct reports**, Agrawal reports shipping **2–10 PRs a week** by combining short steering blocks with one overnight agent run. He reserves the work for prototypes, internal tools, codebase gardening, and product vision.

Gather business and technical context before the run, ask for tests first, exercise the flow with computer use, and have a clean subagent review the result. In the morning, test it yourself, read every change, and split the work into small, reviewer-friendly PRs.

Do not put manager-owned experiments on the critical path or let prototypes reach production directly. Agrawal says his own agent-assisted code has annoyed engineers and caused incidents, so CI, feature flags, a prototype branch, and human review remain necessary.
Connected Context · Feed7 Judgment

This turns agent-assisted prototyping into a bounded leadership practice rather than a general production-delivery model. It confirms the prior emphasis on tests, CI, review, and small PRs, while sharply narrowing autonomy to noncritical experiments and codebase maintenance. The reported incidents add practitioner evidence that overnight throughput does not remove engineering ownership or justify placing agent work on the critical path.

Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI LabBoth treat agent coding as a governed team system; Agrawal adds a concrete manager workflow and direct evidence that missing safeguards can create incidents.Loop Engineering from First Principles — Kyle Mistele, HumanLayerThe instruction to split overnight output into small, reviewer-friendly PRs implements the candidate’s bounded-change and human-pause principle.Effective harnesses for long-running agentsBoth structure long-running work around tests and human verification, but Agrawal narrows the pattern to one overnight run and noncritical tasks rather than multi-session feature delivery.How to Kill the Code Review — Ankit Jain, AviatorAgrawal’s clean-agent review and mandatory human reading reinforce evidence-oriented review, while his incidents caution against treating automated review as sufficient.
Context Map
agentcoding#coding-agents#harness-engineering#skills
Uncertainty
Do not put manager-owned experiments on the critical path or let prototypes reach production directly. Agrawal says his own agent-assisted code has annoyed engineers and caused incidents, so CI, feature flags, a prototype branch, and human review remain necessary.