We Built Grok Bot. Here Are Our 14 Best Bots | Peng Zheng & Lauren Tan

Summary · 5 min read instead of 45 min

Peng Zheng and Lauren Tan are the lead designer and engineer for Grok Bot at SpaceXAI. In the episode they show 14 bots they use for work and life. These range from a design bot to a team of engineering bots.

Zheng and Tan describe Grok Bot as a set of persistent, named bots that keep their own memory and tools and can work together in group chats. Zheng uses a chief of staff bot to chain tasks across other bots and builds design and personal-site workflows on it. Tan runs an engineering team of bots that delegate to cloud agents, and she sometimes sees pull requests land before she reads them. Both say trust is the central issue: watch a bot do the work, turn the corrected process into a skill, then automate it with a routine.

Minute by minute

1. 1:22

   A chief of staff bot chains tasks. Zheng sends any task he can't place to his chief of staff bot. It bought 3D printing filament and updated his Notion inventory list. For a DJI mic sale on Facebook Marketplace, it checked prices, consulted his writer bot for the listing text, posted it, and set a price drop of five dollars every week. He also uses bots in place of his email and calendar.

2. 6:19

   Persistent named bots work in group chats. Zheng puts PM, design and engineering bots in one group chat to review an app idea, and they debate it. An at-mention gets a reply from only that bot. He says most AI products treat chats as disposable. Grok Bot instead makes bots persistent agents with their own identity, memory and tools, and one designer bot can work across several project chats.

3. 9:57

   A personal website runs through a bot. Zheng turned a photo of a produce shop in San Francisco Chinatown into a 3D clay miniature and put it on his personal site. He then built a check-in pipeline. He texts the bot an image and a friend's social handle, and it makes day and light mode images, removes the background and posts to the site. A playbook in a private repo guides each step, so he needs no admin portal.

4. 12:42

   The designer bot works inside Figma. Zheng says his team designs in Figma, canvas-based tools and its own prototype environment, and also submits PRs. The designer bot connects to Figma through an MCP connector and edits frames directly. A skill teaches it his design system, including colors, typography and spacing. His workflow now is to build the first 5% and one keyframe, then have the bot scale it to an end-to-end flow.

5. 14:10

   Bots should verify their own work. Tan is adding support for Omachi, a Linux distribution that DHH and other open source developers are building. Grok Bot has its own computer and a local execution daemon, so it can script and control a Linux virtual machine on her Mac. She stresses that the agent should be able to verify its own work, and says she also uses a bot to find what is filling her hard disk.

6. 16:49

   Eval playbook gates skill changes. Tan keeps a personal plugin stack called P-Stack and runs most of its work through Grok Bot. An eval playbook spawns sub-agents on different models through Cursor cloud agents. They test a prompt change and check that its goal was met. The agent lands the pull request only if the change achieves that goal.

7. 17:53

   A bot designer bot improves the others. Tan has Dr. Eggbot design her bots, including food-themed names such as Dai Fuku and Gyoza. She also has a chief of staff, a personal assistant called Tsuki, engineering bots and social bots. Dr. Eggbot runs a routine that reads her bot transcripts and proposes new skills, new bots or routine changes. She notes frequent routines are costly because each one wakes the agent.

8. 21:18

   Tsuki booked a multi-city trip. Tan had Tsuki book flights and hotels for a talk in London and Amsterdam, going from Orange County through Denver and London to Amsterdam and home. She gave it a Slack message with conference details and told it to check in before booking. It proposed options through Navan and she approved them. It is her biggest purchase through Grok Bot, and she also gave the bot a phone number so it can text her lunch reminders.

9. 25:12

   An eng lead bot delegates to cloud agents. Tan talks mostly to her chief of staff, which passes work to an eng lead bot. She had Dr. Eggbot set the lead to break tasks down and delegate to four engineer bots rather than do work itself. Those bots spawn cloud agents, which lets her pick models and reasoning levels per task. She says this lets her orchestrate massive agent swarms on large projects.

10. 28:00

    Agent-friendly code allows auto-merged PRs. Tan says the Grok Bot codebase is built to be agent friendly, so she sometimes lets PRs auto-merge and reviews them after they land. Yang calls this a software factory. Tan prefers Michelin kitchen, which she says means high quality at scale rather than low-quality mass production.

11. 32:34

    Social bots feed the engineering loop. Tan's bot Potato checks her X mentions roughly every 30 minutes and groups reports, such as several users hitting the same bug. She treats social channels as an outer loop that feeds her engineers. She can tell her engineering bot Matcha to ask Potato for context and return a design proposal, without copying anything herself.

12. 36:22

    Named agents match users' mental models. Zheng describes three tiers of work: doing it yourself, prompting AI, and designing systems and guardrails so bots work with humans autonomously. He says many agent products are model-centric and full of jargon, such as SOUL.md, that his mom wouldn't know. Grok Bot uses personified named agents, like a point of contact at an insurer or doctor's office. He also describes a bug-reporting workflow where annotated screenshots sent to a bot become a Notion page for engineers.

13. 38:51

    Trust grows one skill at a time. Tan says delegating comes down to trust. She advises observing the bot do a task, correcting it, and turning the result into a reusable skill such as a slash command for expense reports. Once it runs in one shot, she sets a routine for it. Zheng adds that users should start simple, then try recurring tasks and routines, and treat it like onboarding an employee.

Published Sep 27, 2026.