Grok Bot: Treating AI Agents as Colleagues, Not Software Tools

Paradigm Shift: Colleagues, Not Tools

Current agent runtimes like WorkBuddy and Codex treat agents as software requiring a management dashboard — you create, configure, assign tasks, and inspect results. Grok Bot, built by the Cursor team after its acquisition by Musk and integration into SpaceXAI, proposes a different model: treat agents as colleagues. A colleague has its own computer, uses your everyday tools, works independently, communicates via instant messaging, and continues running after you close your laptop.

Kevin and Roshan from the SpaceXAI product team shared their design manifesto: deliver an AI team that ships real work, not a better agent control panel. The difference is profound: a dashboard agent needs you to assign tasks, track progress, and check results; a colleague receives a goal, decomposes it, collaborates with peers, and returns only when done or blocked.

A Real Bot Team Configuration

Each Bot is a named, persistent AI teammate with its own cloud computer (browser, filesystem, terminal) logged into your tools. Multiple Bots share one cloud computer — files, browser sessions, app logins — enabling seamless handoff without reconfiguration. Each Bot has independent screen, memory, and context for parallel work. Security boundaries sit at the account level, not per Bot.

Unlike chat assistants that leave the last 10% for you to copy-paste, Grok Bot results land directly in your actual tools.

Why Multiple Specialized Bots?

Attributability: Data issues go to Ashley, design to Pixel. Roshan noted a single agent overloads his mental capacity.

Parallelism: Multiple Bots work simultaneously without blocking each other.

Scoped Memory: Each Bot learns within its domain. Chief of Staff shouldn't debug evals; eval Bot shouldn't archive emails.

Mirrors Real Organizations: Companies comprise specialists. One Bot per role; start with one, expand gradually.

Kevin's Team Bot Roster

Kora, Chief of Staff: Generalist managing calendar, Slack, inbox; maintains an "Attention List" that stays quiet when nothing changes.

Emily, Engineering Manager: Explicitly trained not to write code. Decomposes PRDs into technical tasks, dispatches to engineer Bots, accepts deliverables against goals.

5 Engineer Bots: Receive Emily's tasks, invoke Cloud Agents for coding, coordinate and cross-verify.

Ashley, Data Analyst: Accesses data lake and user research library; delivers daily dashboard insights in team-preferred chart styles.

PM Pete, Product Manager: Writes RFCs/PRDs, conducts research, aggregates feedback, coordinates across Bots; collaborated with Kevin on Raspberry Pi hardware.

Pixel, Designer: Works in Figma and design systems.

Rey, Recruiting: Sources talent, drives interview pipelines.

The mapping is direct: human organizational roles become Bots with clear responsibility boundaries, delivery standards, and collaboration links.

Four Core Workflows

Attention List: A New Work Primitive

Not a priority list you write, but an emergent view inferred from your actual behavior — Slack replies, email handling, Notion edits. Two uses:

Filter: From thousands of daily messages, surface only relevant items; auto-archive the rest.

Diff: Compare stated priorities vs. actual time allocation to reveal misalignment.

Setup is one natural-language instruction: "Every hour check email, Slack, Granola, calendar; generate a list of projects I'm focusing on, current status, and next steps." The diff often exposes uncomfortable gaps.

Routines: From Reactive to Proactive

A single phrase — "clean my inbox hourly" — made Kora create a scheduled job. Bots detect repetitive actions and propose or create routines. They can learn a workflow by watching you perform it once on their computer and codify it as a routine.

Critical distinction: a skill is "how to do it"; a routine is "who does it and when." Called "Teach a task" — a once-used, never-go-back feature. After X MCP launched, Ashley was tasked to push hourly install-rate pulse charts, replacing manual dashboard refreshing.

Research to PRD: Parallelism, Citations, Bot-to-Bot Collaboration

Live demo using "bidirectional voice mode" (audience-voted requirement):

PM Pete scraped Reddit, X, internal Slack feedback channels, user research library; simultaneously evaluated speech-to-speech model benchmarks for tech selection.

Ashley ran in parallel: market sizing and current voice feature weekly-active penetration, outputting brand-styled charts.

Pete wrote into Notion RFC using team PRD template: opportunity, user pain, solution, TLDR, out-of-scope.

Key detail: user told Pete "loop in Pixel for design." Pete packaged context (seed copy, visual brief) and sent to Pixel; Pixel returned Figma outputs directly into RFC mock placeholders. No human context shuttling. Bots hand off to each other.

Two emphasized practices: require Bots to cite sources for verification; Agents excel at writing prompts for other Agents.

Delivery: Group Chat + Cloud Agents

Pete and Emily added to a group chat: "Pete wrote the PRD, you two sync, Emily start building." Emily confirmed requirements source and V1/MVP scope with Pete, broke PRD into tickets, assigned to engineer Bots, each spinning up Cloud Agents to code. Environments preloaded with codebase, dependencies, secrets, test capabilities. On login/approval needs, Emily paused and @-mentioned humans to unblock.

This "dispatch-and-review" pattern wasn't a one-off command — Emily learned it from past work. Internally, double-digit percentage of merged PRs now originate from Grok Bot-launched Cloud Agents.

Engineering Perspective: A Developer's Real Setup

SpaceXAI engineers use Grok Bot to build Grok Bot — a more aggressive, telling configuration.

Five engineer Bots, each owning a domain: Baltata (iOS shared layer), Shaoruru (desktop client & CI/CD), Hogan (infrastructure), Craig (Android), Quill (harness). Each can work cross-domain but is sharpest in its own area due to more precise specs and design principles.

Engineer Bots don't write code; they manage Cloud Agents. On receiving a task, a Bot launches a Cloud Agent equipped with personal skills (e.g., /lingxi-design for visual, /react-native-best-practices for code quality audit, /lingxi-review for architecture judgment, /lingxi-product for product decisions), plus a detailed prompt specifying deliverables and required evidence. The Bot monitors the Cloud Agent's transcript and screenshots, notifies on completion, proactively unblocks when stuck.

Key mechanism: Cloud Agents can screenshot; Grok Bot's multimodal capability verifies visual changes against expectations, rejecting mismatches. This feedback loop means humans don't stare at screens waiting for results.

Additional role: Jenny, Ops Lead — the only non-coding Bot. Daily 5 AM 1:1s with each engineer Bot to review playbooks, surface blockers, reinforce desired work style. When a Bot errs, Jenny runs root-cause analysis and postmortem, updates playbooks, notifies all — ensuring the same mistake never recurs twice.

Metrics: previously manually managing 15 Cloud Agents; now Bot team manages 200+ concurrently. Colleague poteto submitted 2000+ PRs in one month. Balta and Shaoruru built Grok Bot's infrastructure in four weeks; iOS v0 shipped in three weeks.

Two Notable Extensions

Nightly Audits: 3 AM daily, engineer Bots sweep codebase: remove dead code, optimize load speed, reduce bundle size. Morning brings clean PRs ready to merge. Maintenance shifts from "when remembered" to daily habit.

P0 Urgency Process: Say "this is P0" and Bots spin up a temporary routine checking transcripts every five minutes, actively correcting Cloud Agent drift. Saves huge time but burns tokens fast — reserved for true emergencies.

Security, Platform, Pricing

Passwords, 2FA, CAPTCHA, payment confirmations require manual takeover. API keys use secure secret cards, never exposed in chat.

Practical tips: use Always Allow cautiously; lock destructive actions (delete, pay, send externally) first; start with read-only, expand as needed; give Bots explicit role descriptions and boundaries — they'll find their own limits.

Platforms: macOS, Windows, iOS; Android coming. Desktop and mobile switch seamlessly; same Bot accessible on both.

Pricing tied to existing subscriptions: Cursor Pro $20/mo, SuperGrok $30/mo, Teams $40/seat/mo. Grok Bot usage metered separately, not mixed with Grok/Cursor quotas.

Getting started: build one Bot for one thing. Don't assemble a full team day one. Mantra: "One Bot, one job; treat it as a person, as a role." Official TemplateBot marketplace for instant cloning; usegrokbot.com for usage patterns. Quick replication: point any Bot at Kevin's X post — it auto-builds the same team and asks about your work habits.

The Human's Place

Boundary is sharp.

Offloaded to Bots: grunt work and low-complexity tasks — collection, summarization, first drafts, dispatch.

Humans keep the last mile of review and refinement: judgment on research conclusions, strategic PRD layers, code review, any outbound email, purchasing, deletion.

Humans still send messages and edit docs personally. Two reasons: remove AI flavor; recipients need to know the sender actually thought.

Roshan's summary: "I can initiate and manage far more work than before, so I actually have more time for review and thinking about what's next — exactly where a product person should spend time."

The engineer building Grok Bot with Grok Bot put it bluntly: "I wake up to a stack of PRs waiting for merge. Code quality meets my bar; visuals match my aesthetic. More work passes on first try, letting me focus on harder architecture decisions and performance optimization."

Side note: Tokenbank (a decentralized open-source project for agent/model sharing) considers a similar assumption, but differs — Grok's colleagues use official cloud computers, while Tokenbank's hired colleagues reuse the provider's machines. Benefit: hirer avoids local agent misbehavior and compute cost; provider prevents easy copying of their agents, preserving value. Downside: hirer must hand requirements to external agents, unsuitable for high-privacy scenarios.

Back to the opening contrast. WorkBuddy and Codex give you a better steering wheel. Grok Bot gives you a fleet. Neither is wrong — one serves current user habits, the other designs for the future. The latter transcends efficiency-tool territory; it's a more thorough AI-Native form. It productizes "managing a team" itself. The PM's core skill shifts from hands-on execution to defining goals, configuring teams, reviewing outcomes. The engineer's core skill shifts from writing code to defining standards, managing Bot teams, making final architecture calls.