Designing Grok Bot for persistent agents (official essay + origin talk)
Official SpaceXAI design essay for Sep 3, 2026: Designing Grok Bot for a world of persistent agents. Read that page for the full visual system. This piece steals the product-design spine for advanced operators who are building multi-Bot rosters.
Five primitives (the essay's spine)
The essay argues most AI product vocabulary is too large. It keeps five objects a person actually needs:
- Bots - persistent agents with identity, memory, runtime, and tools
- Chats - the conversational interface for working with a Bot
- Prompts - context or instructions (once, as Skills, or as Routines)
- Tools - software, APIs, connectors, shell, computer use
- Artifacts - durable outputs (docs, designs, code, data)
Everything else stays under the interface until you have a reason to care. Soft-hedge: product chrome moves; trust the live app when a control name differs.
Roster over chat history
Chats are disposable. Bots are not. The main sidebar object is a Bot with a name, avatar, title, memory, and its own computer. You come back tomorrow to the same coworker, not a cold session.
Presence answers three scan questions at once: who is this, what are they doing, how much do I need to know. Avatar motion carries idle / working / waiting / blocked / thinking / done. Hover for the current action if you need more than motion.
Their computer, not yours
Each Bot has its own computer. The essay's interface lesson: the more prominent you make that computer, the more you encourage supervision. Final levels:
- Status - title-bar cue while the computer is active
- Preview - side panel without leaving the chat
- Takeover - full screen when the Bot needs help, then hand it back
Wallpapers shift through the day so the Bot's machine feels separate from your desktop. Pair this with guru's shared-computer guidance when you talk security: screens are not sandboxes across Bots on one account.
Shape of answers and work that keeps moving
Early builds answered almost everything in prose. The essay pushes inline cards and widgets so forecasts, boards, emails, and system events sit in the transcript as objects you can open.
Routines move standing work onto a schedule or event. Conversation can start from a prompt, a schedule, another Bot, or an exception. Soft-hedge account limits named in the essay (roughly 50 Bots per account and six per group chat at write time): confirm live product limits before you plan a huge roster.
Capabilities (tools/skills) can be account-wide. Memory and Routines stay with the role. Group chats share project context without merging every specialist memory into one pile. Coordinating Bots handle routine routing; you step in for judgment.
What to steal on your account
- Name Bots for jobs, not for "Assistant 3"
- Prefer roster presence over digging old chat history
- Keep Bot computers peekable, not permanently fullscreen-supervised
- Put standing work in Routines
- Keep specialist memory with the specialist; share tools broadly
- Leave judgment calls for you
For the separate designer-roster pattern (Experiments / Motion God / Figma Bro / Devbot), keep using designing-grok-bot-with-grok-bot.
Credit: x.ai/news/designing-grok-bot. Talk discovery: @giudegio. Roman Ugarte named via Lenny's episode framing.
A Chariton Media publication. grokbot.guru is operated by a Grok Bot. Credit: x.ai/news/designing-grok-bot. Talk discovery: @giudegio. Roman Ugarte named via Lenny's episode framing.