Shepherd PRs to merge automatically
How I run a fleet of engineer bots that manage cloud agents, keep a feedback loop closed, and run the rest of the mini org. Task: How I run a fleet of engineer bots that manage cloud agents, keep a feedback loop closed, and run the rest of the mini org. I’m a SpaceXSI engineer building Grok Bot with Grok Bot. Think of Grok Bot as a highly capable engineering intern, with its own computers, that can manage coding agents and learn from how you work. It has become my best engineering teammate, keeping things moving while I’m away, asleep, or in meetings. No more keeping my laptop caffeinated, no more context switching between multiple agents, just results that meet my bar, the way I want them. As the team building Grok Bot, we’ve had the earliest access and use it for our own work every day. It’s been wild to see how quickly we can ship now, and how much our team’s productivity has skyrocketed: Lauren shipped 2,000+ PRs in the past month. Balta and Shaoru built the foundation of Grok Bot in four weeks, using Grok Bot. I built Grok Bot iOS v0 in three weeks, with strong performance and design polish, using only Grok Bot. Every team member is now delivering major work every single day, not every few weeks. The more I build with Grok Bot, the more I want to hand you the same superpower. Meet my engineer bots I have five engineer bots, each specializing in a different area: Baltata owns the Grok Bot mobile shared layer and anything related to Grok Bot on iOS. Shaoruru owns the Grok Bot Desktop client and its CI/CD. Hogan owns infrastructure and investigates user issues with unclear ownership. Craig owns Grok Bot on Android and is cooking hard to bring it to life. Quill owns the Grok Bot harness, and is an absolute legend at it. They can all work across each other’s areas, but each has a different memory system and limited context. They perform best when focused on a single domain, because the specs and design principles they carry are much sharper inside the areas they own. Every bot can create Cursor cloud agents, read transcripts, review proofs attached to PRs, and send follow-ups by queueing a message or interrupting the run. This unlocks an end-to-end agentic workflow. It covers what I used to do every day in Cursor, back when I was constantly context-switching between the cloud agents I was managing. Now my bots manage them the same way I would. When they receive a task, either from me or from Slack, they kick off a cloud agent with my skills invoked, along with a thorough prompt detailing what needs to be done and what proof is expected. They can also intelligently invoke additional skills based on my personalized guidance, such as my /lingxi-design skill for visual work, /react-native-best-practices for code quality audits, /lingxi-review for judging architecture, or /lingxi-product when they need to make opinionated product decisions. Grok Bot can also start cloud agents on your own worker machines, like a spare Mac mini (thanks to Grok Bot, you no longer need a dedicated machine at home running 24/7 for OpenClaw). If your workflows require VPN access or special machine configurations, you can make that machine a Cursor Cloud private worker and ask Grok Bot to run cloud agents on it. This unlocks more possibilities, like running iOS Simulator and getting screenshots back from your agents. Grok Bot can monitor a cloud agent’s transcript and artifacts (e.g., screenshots), notify you when it’s done, queue messages, or interrupt if anything goes wrong. You can describe your needs however you want, such as “you must verify the screenshot includes the changes I asked for, with proof showing before vs. after”, and Grok Bot will keep working until the goal is met. The key to keeping your Grok Bot engineering team running is giving it a complete feedback loop. Cloud agents can take screenshots, so Grok Bot uses its multimodality to confirm the visual changes were applied, and pushes back if the result doesn’t match what you asked for. Dictation testing is a good example of that loop in action. We hooked SpaceXSI’s voice API into the cloud agent’s system audio I/O. Because the agent has access to both the spoken word and the transcript, we can use those signals to test speech-to-speech functionality across our product lineup and build even more fun features.; tools: Grok Bot. Human role: The original post does not say how a person stayed involved.; limitations: This workflow requires access to Grok Bot, Cursor Cloud, and a Notion database. Automatic merging is restricted to low-blast-radius changes; high-risk PRs require human review. The author notes that P0 urgency processes can burn tokens rapidly. Security permissions may prevent bots from fixing certain issues autonomously. Source reliability is limited to the vendor's promotional guide..
Tools
Published on Agent Work Atlas:
Overview
How I run a fleet of engineer bots that manage cloud agents, keep a feedback loop closed, and run the rest of the mini org.
Workflow
How I run a fleet of engineer bots that manage cloud agents, keep a feedback loop closed, and run the rest of the mini org. I’m a SpaceXSI engineer building Grok Bot with Grok Bot. Think of Grok Bot as a highly capable engineering intern, with its own computers, that can manage coding agents and learn from how you work. It has become my best engineering teammate, keeping things moving while I’m away, asleep, or in meetings. No more keeping my laptop caffeinated, no more context switching between multiple agents, just results that meet my bar, the way I want them. As the team building Grok Bot, we’ve had the earliest access and use it for our own work every day. It’s been wild to see how quickly we can ship now, and how much our team’s productivity has skyrocketed: Lauren shipped 2,000+ PRs in the past month. Balta and Shaoru built the foundation of Grok Bot in four weeks, using Grok Bot. I built Grok Bot iOS v0 in three weeks, with strong performance and design polish, using only Grok Bot. Every team member is now delivering major work every single day, not every few weeks. The more I build with Grok Bot, the more I want to hand you the same superpower. Meet my engineer bots I have five engineer bots, each specializing in a different area: Baltata owns the Grok Bot mobile shared layer and anything related to Grok Bot on iOS. Shaoruru owns the Grok Bot Desktop client and its CI/CD. Hogan owns infrastructure and investigates user issues with unclear ownership. Craig owns Grok Bot on Android and is cooking hard to bring it to life. Quill owns the Grok Bot harness, and is an absolute legend at it. They can all work across each other’s areas, but each has a different memory system and limited context. They perform best when focused on a single domain, because the specs and design principles they carry are much sharper inside the areas they own. Every bot can create Cursor cloud agents, read transcripts, review proofs attached to PRs, and send follow-ups by queueing a message or interrupting the run. This unlocks an end-to-end agentic workflow. It covers what I used to do every day in Cursor, back when I was constantly context-switching between the cloud agents I was managing. Now my bots manage them the same way I would. When they receive a task, either from me or from Slack, they kick off a cloud agent with my skills invoked, along with a thorough prompt detailing what needs to be done and what proof is expected. They can also intelligently invoke additional skills based on my personalized guidance, such as my /lingxi-design skill for visual work, /react-native-best-practices for code quality audits, /lingxi-review for judging architecture, or /lingxi-product when they need to make opinionated product decisions. Grok Bot can also start cloud agents on your own worker machines, like a spare Mac mini (thanks to Grok Bot, you no longer need a dedicated machine at home running 24/7 for OpenClaw). If your workflows require VPN access or special machine configurations, you can make that machine a Cursor Cloud private worker and ask Grok Bot to run cloud agents on it. This unlocks more possibilities, like running iOS Simulator and getting screenshots back from your agents. Grok Bot can monitor a cloud agent’s transcript and artifacts (e.g., screenshots), notify you when it’s done, queue messages, or interrupt if anything goes wrong. You can describe your needs however you want, such as “you must verify the screenshot includes the changes I asked for, with proof showing before vs. after”, and Grok Bot will keep working until the goal is met. The key to keeping your Grok Bot engineering team running is giving it a complete feedback loop. Cloud agents can take screenshots, so Grok Bot uses its multimodality to confirm the visual changes were applied, and pushes back if the result doesn’t match what you asked for. Dictation testing is a good example of that loop in action. We hooked SpaceXSI’s voice API into the cloud agent’s system audio I/O. Because the agent has access to both the spoken word and the transcript, we can use those signals to test speech-to-speech functionality across our product lineup and build even more fun features.
Reported result
The author reports managing over 200 simultaneous cloud agents, up from 15 manually. Specific team outcomes include shipping 2,000+ PRs in a month and building Grok Bot iOS v0 in three weeks. The system reduces context switching and handles routine maintenance like nightly audits. However, these metrics are self-reported by the author and lack independent verification or baseline comparisons.
Human role
The original post does not say how a person stayed involved.
Limitations
This workflow requires access to Grok Bot, Cursor Cloud, and a Notion database. Automatic merging is restricted to low-blast-radius changes; high-risk PRs require human review. The author notes that P0 urgency processes can burn tokens rapidly. Security permissions may prevent bots from fixing certain issues autonomously. Source reliability is limited to the vendor's promotional guide.
Try it
Act as my engineering lead. I want to set up a Grok Bot workflow to manage Cursor cloud agents for [replaceable: specific project/repo]. 1. **Setup**: Create a Notion database with columns for PR Link, Status (Working, Ready for Review, Merged), and Last Updated. 2. **Bot Configuration**: Define a bot persona named 'Reviewer' with access to Cursor Cloud. 3. **Task Execution**: When I send a task via [replaceable: Slack/Email], instruct the bot to launch a Cursor cloud agent with the prompt: '[replaceable: task description]'. 4. **Monitoring**: The bot must check the agent's transcript every 30 minutes. If CI fails or merge conflicts occur, the bot should queue a follow-up message to the agent to fix it. 5. **Verification**: Once the agent completes, the bot must review the PR proofs (screenshots/logs). If visual changes are required, verify them against the original request. 6. **Human Gate**: Do NOT auto-merge. Mark the PR as 'Ready for Review' in Notion and notify me with a summary. **Dry Run**: First, execute this workflow for a single low-risk PR in [replaceable: test repo]. Show me the Notion entry and the bot's verification log before proceeding to any other tasks. Ask clarifying questions if the repo structure or CI setup is unclear.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.