The AI Builder Podcast
How a Google AI PM Runs an Agent Team for Research, GitHub and Content
Shubham Saboo runs a scheduled agent squad — chief of staff, research, triage, content — that supports a 110,000-star GitHub repo and escalates only what needs a human.
What you’ll learn
- The squad: a chief-of-staff agent, research agent, engineering-triage agent, platform-specific content agents and a newsletter agent — all managed through Telegram.
- Cron jobs and daily decision queues make agents proactive; humans only see the decisions that need them.
- 30-day performance data, self-review loops and biweekly squad reviews keep the agents accountable.
- OpenClaw is the squad layer; Hermes is the operating interface underneath.
- Start with one repetitive workflow, give the agent context and boundaries like a new hire, and expand only after it performs reliably.
About this episode
Shubham Saboo supports a 110,000+ star GitHub repository and content for more than 350,000 developers — with a scheduled team of AI agents doing the recurring work.
Managed through Telegram, the squad includes a chief-of-staff agent, a research agent, an engineering-triage agent, platform-specific content agents and a newsletter agent. Cron jobs and daily decision queues make the agents proactive rather than reactive; 30-day performance data, self-review loops and biweekly squad reviews keep them accountable; and only the decisions that genuinely need a human ever reach one. Underneath, OpenClaw is the squad layer and Hermes the operating interface.
His starting advice is deliberately unglamorous: automate one repetitive workflow, give the agent context and boundaries the way you would a new hire, and expand its responsibilities only after it performs reliably.
Questions this episode answers
What agents are in Shubham Saboo's squad?
A chief-of-staff agent, a research agent, an engineering-triage agent, platform-specific content agents and a newsletter agent — all managed through Telegram, running on cron schedules.
How do the agents stay accountable?
Through 30-day performance data, self-review loops and biweekly squad reviews. Daily decision queues surface only the calls that need a human.
How should you start building an agent team?
Automate one repetitive workflow first. Give the agent context and boundaries like a new hire, and expand its responsibilities only after it performs reliably.
