We recently sat down with Lauren Tan (Potato on X) for an AMA. As the creator of PStack and a key mind behind GrokBot, Lauren talked through how to orchestrate AI agents, build trustworthy codebases, and what is next for the GrokBot ecosystem.
GrokBot as the "outer loop"
A common question for developers is how GrokBot fits into a workflow that already includes tools like Cursor. Lauren frames the two as complementary:
- The outer loop: GrokBot is strong at broad connections — Slack, calendars, email — to manage routines, watch for feedback, and trigger workflows.
- The inner loop: Cursor stays the place for heavy, deep-dive engineering.
- Specialized teams: One persistent bot can work, but a team of specialized bots, each with its own isolated memory and context, is a more effective way to organize projects.
The power of agent-written skills
The elaborate instructions inside PStack are intentionally written by AI, not by hand.
- Continuous evaluation: When agents write their own operational prose, they can independently evaluate, score, and iterate on their skills in a loop.
- Defaulting to rigor: It is hard to know how complex a bug is up front. Defaulting to high-reasoning modes (like
/potato mode) makes agents verify and reproduce issues before they write code.
Building trust with architectural constraints
Lauren flagged a real risk of AI-assisted coding: agents are essentially copying machines. If the codebase is full of legacy anti-patterns, they will replicate that technical debt like a virus.
- Strict frameworks: To automate pull requests with confidence, Lauren’s team built "Dune," an internal React framework with highly pedantic lint rules.
- Categorical impossibility: Restrict the APIs and lock the architecture (for example, banning
useEffect) and it becomes impossible for an agent to merge bad code. - Gardening strategy: The safest refactor is to write a lint rule against a bad pattern, have the agent fix every existing violation, and merge it all at once so the AI has no bad examples left to copy.
What’s next for GrokBot
The roadmap points toward more capability and more natural interaction:
- A new version of the Grok model is on the horizon, with better performance and capability.
- The team is experimenting with channels and threads so different bots can talk to each other directly.
- Voice mode is in iteration, aiming for a conversational, real-time interface with your bots.