Building Grok Bot with Grok Bot: Scaling AI Agents to 200+ Parallel Workers
SpaceXAI engineers use Grok Bot, a system of specialized AI agents that manage Cursor cloud agents, automate code reviews, run nightly audits, and scale to 200+ parallel workers, delivering 2000+ PRs monthly while eliminating context switching and enabling continuous delivery during off-hours.
Introduction
Grok Bot functions as a highly capable engineering intern: it has its own computer, manages coding agents, and learns from the user's work style. It continuously pushes work forward while the user is away, asleep, or in meetings, eliminating the need to keep a laptop alive and removing context switching between multiple agents. The output meets the user's standards and is done the way they want.
As the team building Grok Bot, early access and daily use yielded striking productivity gains:
@poteto submitted 2,000+ PRs in the past month.
@baltaaazr and @shaoruu built the Grok Bot foundation in four weeks using Grok Bot.
The author built Grok Bot iOS v0 in three weeks using only Grok Bot, with strong performance and polished design.
Now every team member delivers meaningful work daily instead of once every few weeks.
Engineer Bots
Five specialized engineer bots each own a distinct domain:
Baltata — owns the Grok Bot mobile shared layer and everything iOS-related for Grok Bot.
Shaoruru — owns the Grok Bot desktop client and CI/CD work.
Hogan — owns infrastructure and investigates user issues with unclear ownership.
Craig — owns Grok Bot on Android, pushing hard to ship it.
Quill — owns the Grok Bot harness.
They can collaborate across domains but each has its own memory system and limited context. They perform best when focused on a single domain because the specs and design principles they carry are much clearer within their own scope.
Each bot can create Cursor cloud agents, read transcripts, view artifacts attached to PRs, and follow up by queuing messages or interrupting runs. This unlocks an end-to-end agent workflow covering what the author used to do daily in Cursor — constantly switching context between cloud agents managed manually. Now the bots manage those agents the same way.
When they receive a task — whether directly from the user or from Slack — they spin up a cloud agent, invoke the user's skills, and attach a detailed prompt describing what to do and what proof is required. They can also intelligently call additional skills based on personalized guidance: /lingxi-design for visual work, /react-native-best-practices for code-quality reviews, /lingxi-review for architectural judgment, or /lingxi-product for opinionated product decisions.
Your Mac Mini Is Now a Cursor Agent
Grok Bot can also launch cloud agents on the user's own machine — for example, an idle Mac mini (eliminating the need for a dedicated 24/7 machine running OpenClaw at home). If the workflow requires VPN access or special machine configuration, that machine can be designated as a Cursor Cloud private worker and Grok Bot will run cloud agents on it. This unlocks possibilities like running the iOS Simulator and retrieving screenshots from agents.
Closing the Feedback Loop
Grok Bot monitors cloud agent transcripts and artifacts (e.g., screenshots), notifies on completion, queues messages, or interrupts when issues arise. Requirements can be described in any way — for instance, "You must verify the screenshot contains the changes I asked for and provide before/after proof" — and Grok Bot will keep working until the goal is met.
The key to keeping the Grok Bot engineering team running is giving it a complete feedback loop. Cloud agents can take screenshots, so Grok Bot uses its multimodal capability to confirm visual changes have been applied; if the result doesn't match the request, it pushes back.
Dictation testing exemplifies this loop: SpaceXAI's voice API was wired into the cloud agent's system audio I/O. Because the agent has access to both spoken word and transcript, those signals are used to test voice-to-voice features across the product line and build more capabilities.
Sometimes agents hit environment flakiness and stall until a follow-up prompt. Grok Bot handles this: it continuously watches runs and unblocks agents as aggressively as possible. Since adopting Grok Bot, one-off environment flakiness almost never reaches the user; the exception is when Grok Bot lacks permissions to fix it safely.
Everything is now one message away — e.g., asking agents to iterate 10 more times before handing off.
Scaling Beyond Context Limits
To help bots keep up after exceeding context limits — and avoid scrolling huge chat histories — each engineer bot manages a shared Notion database.
Every 30 minutes they scan the database and check each PR for:
Bugbot comments or security findings, verifying each one.
Failed CI runs.
Merge conflicts.
If anything is off, they immediately follow up with the cloud agent and move that row back to "Working". If everything looks clean, they mark the task "Ready for Review" and automatically trigger a code review focused on quality and potential blind spots.
If the review is high-confidence and small in scope, the PR auto-merges. Otherwise, the user reviews code and proofs and decides to merge or give feedback.
Almost every morning tasks are ready to merge. Code quality meets the bar, visuals match intent, and proofs clearly show what was tested. More work gets done in one shot, freeing focus on harder problems, higher performance bars, finer visual polish, and bigger architectural decisions.
Before Grok Bot the author could manually manage 15 cloud agents at once. Now the agent fleet manages 200+ simultaneously, with further scaling possible.
Grok Bot Runs This Mini-Org
Beyond engineering, operational overhead includes onboarding new engineer bots, sharing knowledge, running postmortems (e.g., a PR not properly checked), and daily standups for alignment. All of this is handled by Jenny, the ops lead and the only agent who doesn't write code.
Every morning at 5 AM, Jenny holds one-on-ones with every agent, reviewing the playbook, surfacing blockers, and reinforcing the desired culture. This has proven remarkably effective — even after many weeks, agents rarely forget complex workflows.
When a bot makes a mistake (e.g., not pushing hard enough until the real goal is clarified), it is sent to Jenny for root-cause analysis and a postmortem. Jenny digs into the reasoning that led to the issue, updates the playbook, and notifies the other engineer bots so the same mistake doesn't recur.
When the team needs to grow, Jenny handles onboarding: she creates the new bot in the organization, shares engineering team rules, and enlists Hogan and the rest of the team to help with ramp-up.
The goal of the full engineering system inside Grok Bot is to minimize repetitive work. Delegate to Grok Bot to focus on harder, deeper problems.
Bonus Use Cases
Nightly Audits
Every day at 3 AM engineer agents clean the codebase, improve code quality, sweep dead logic, speed up app launch, and reduce bundle size. Every morning a fresh batch of PRs keeps the code clean, cruft-free, and scalable. Code maintenance becomes a daily habit instead of an occasional event.
More nightly audit ideas:
Security audits to catch issues the team might miss.
CI/CD build-time audits to prevent build times from creeping up unhealthily.
Internationalization audits to fill gaps when features ship in only one language.
Consistency audits to avoid drift when building multiple clients (iOS vs. desktop) and features land on only one side.
Catch-up audits that monitor PRs merged in the last 24 hours in areas of interest, returning a high-level summary and a curated list of PRs to review.
Favorite prompt: "You have six hours tonight. Do whatever you want. Have fun!"
P0 Emergency Flow
Cloud agents can be slow — they need to spin up, set up environments, wait, run tests, and iterate. For urgent needs, a P0 emergency flow was designed: whenever a task is labeled P0, an ad-hoc routine checks transcripts every five minutes, monitors progress and reasoning, and actively steers the cloud agent when it starts wasting time.
This works extremely well for urgent codebase research or critical bug fixes. Note: it burns tokens faster than expected, so use only when truly urgent.
Lessons & Tips
Give cloud agents a full feedback loop: They need signals on what to do next without the user. They should be able to spin a dev instance and drive the full stack end-to-end (e.g., via Chrome DevTools, CLI, or Apple Accessibility). If they can't, let them run the process, proactively unblock themselves as much as possible, and package what they learn into a reusable repo skill.
Treat Grok Bot like a talented intern: If communicating an engineering task is difficult, have it do homework first — learn weak areas, reference how other engineers do the work. No skill invocation, no long prompts, just chat.
Avoiding repetition is key: As AI gets more capable, delegate repetitive tasks and focus energy on deeper, harder problems agents don't easily solve. If you find yourself doing something more than once a day with a clear pattern, discuss with your bot how it can help.
Daily standups with bots work surprisingly well: Repeating key points daily helps them remember complex workflows while juggling many tasks. Since context length can't hold everything, daily reminders are a useful nudge that saves repetition.
Let go more: Collaborating with agents is a trust-building process, like self-driving. Think about when they run smoothly vs. when they might fail. Give them freedom to deliver in safe areas; be more careful in high-risk zones. But don't stop them just because they failed before — keep experimenting and helping them grow.
Let them orchestrate each other: Bots are more capable than you think. To take ops hands-off further, build a bot error-review pipeline (e.g., an ops bot talks to other bots and analyzes their thought traces) so the same mistakes don't repeat.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
