Episode 8 · Hiten Shah · August 28, 2026 · 57 min
YouTube · Spotify · Apple Podcasts
About this episode
Serial founder and Crazy Egg CEO Hiten Shah thinks most people are still using AI like Google: they type a short request, accept the first answer, and stop. His alternative is to give AI real context, show it what good work looks like, and let specialized bots carry a job further—while a person keeps the final judgment.
In this episode, Michael and Hiten compare the friction of earlier agent setups with Grok Bot, which gives each bot a computer it can use. Hiten explains how he connects research, product-marketing, and librarian bots; uses private GitHub repositories as shared context; and limits risky actions with read-only access, explicit constraints, and approval steps.
The examples are concrete. Hiten’s bots made 1,800 GitHub commits in a day before he corrected the behavior and set a limit. In another workflow, a local AI system reproduced a desktop-app bug, gathered screenshots and video, debugged the problem, and prepared a pull request for human review.
What you’ll learn
- Why Hiten calls today’s shallow AI use “the Google problem”
- What changes when a bot has its own computer instead of only a chat box
- How a research bot, product marketer, and librarian can work as a team
- Why Hiten stores bot-created context and artifacts in private GitHub repositories
- How he handles access, permissions, read-only connections, and approval gates
- Why he does not use one chief-of-staff bot as a bottleneck
- How to get better output by showing AI examples of what good looks like
- Why human judgment remains the final mile before anything is published or shipped
- How local models can support QA and move work from a bug report to a reviewed pull request
- The product question Hiten now asks repeatedly: “If we were to build it today, what would we build?”
- Why distribution and go-to-market may need to shape the product from the beginning
Topics and resources
- Hiten Shah
- Crazy Egg
- Typeahead
- Grok Bot
- Cursor
- OpenClaw
- Hermes Agent
- Last 30 Days
- GitHub
- ChatGPT
- Claude
- Codex
- Granola
One experiment to try
Try one low-risk experiment: choose a task you already understand, give AI the relevant context and one strong example, constrain what it can change, and review the result before it acts publicly.