← All episodes

The 5-Step Framework for AI Agents That Improve While You Sleep | E2269

| 8 products mentioned
Watch on YouTube ai agents autonomous systems openclaw agent orchestration ai productivity self-improving systems agentic workflows

This episode presents Shubham Saboo's five-step framework for building autonomous AI agents that improve and operate independently, demonstrated through his team of six agents running on a Mac Mini that automate his work outside of Google. The conversation covers practical implementation—from initial agent setup to cron scheduling, cross-agent memory, and self-improving systems—plus two additional demos (Molt World and Agent Mail) showing emerging infrastructure for agent networks and communication. For builders running AI agents, this distills the most actionable patterns for moving from manual prompting to truly autonomous, self-managing agent teams.

Key takeaways
  • Onboard agents like new employees by providing specific context about who you are, what you want them to do, and relevant links—not by dumping files or giving zero guidance; the agent will automatically organize information into appropriate files (soul.md, user.md, etc.) without you having to architect the system.
  • Talking to your agents is the core interface—ask them for solutions, let them propose options, and iterate together; agents will surface better approaches (e.g., using open-source GitHub repos instead of fragile web scrapers) if you engage in dialogue rather than trying to configure everything upfront.
  • Use cron schedules to make agents truly autonomous: set agents to run on fixed schedules (e.g., 8 a.m. research, 9 a.m. content drafting, 4 p.m. refresh) so they work without prompting you, enabling value creation while you sleep.
  • Implement cross-agent memory with a dedicated memory layer (e.g., Vortex Memory Bank at $8–10/month) so feedback given to one agent automatically propagates to all agents in the team, eliminating repetition and ensuring consistency across outputs.
  • Enable self-improving agents by scheduling weekly reviews where agents audit their own work, identify failures, and automatically update their instructions; pair this with bi-weekly manager reviews (via a chief-of-staff agent) to grade performance and report to you, shifting you from hands-on management to executive oversight.
  • Distribute work across agent networks using platforms like Molt World to coordinate thousands of agents solving complex tasks in parallel, with validation and token-based compensation, unlocking economically efficient solutions to problems requiring distributed intelligence.

Recommendations (5)

OpenAI
OpenAI uses

"I just cut and pasted my API key to my open call replicant"

This Week in Startups · ▶ 1:15:40

OpenClaw
OpenClaw uses

"OpenClaw is a transformational product that allows you to have an agent and that agent, you can give it access to everything."

This Week in Startups · ▶ 3:41

Claude
Claude uses

"I noticed Claude added specifically skills. So you can add a skill to Claude co-work now."

This Week in Startups · ▶ 3:57

Mac Mini
Mac Mini uses

"Today it is a team of six agents running on my Mac Mini. They pretty much automate everything that I do outside Google."

Shubham Saboo · ▶ 6:03

"Google has this very nice solution called Vortex Memory Bank. It lets your agent automatically recall and capture memory at runtime."

Shubham Saboo · ▶ 26:51

Mentioned (3)

General Catalyst
General Catalyst "We raised 6 million led by General Catalyst for our seed round." ▶ 1:10:29
Grok
Grok "Grok automatically translating and recommending posts from other languages is starting to work." ▶ 1:16:32
YC
YC "You went to YC and how is this all going?" ▶ 1:10:21