Cruce Saunders
  • Player card
  • Workshop
  • List
  • Contact
Cruce Saunders
  • Player card
  • Workshop
  • List
  • Contact

Grok Bots

Oct 2026AI agentsPrivate

Flow diagram from the company repo: Cruce sends asks to the Chief of Staff, which sends TASKs with his exact words to a team lead; the lead fans out to subagents and cloud agents, and DONE messages with proof links flow back to one line for Cruce.
How a request moves, from the company repo's own design docs.
Back to the worldGrok Bots
  • Player card
  • Workshop
  • List
  • Contact
  • Back to the world
  • Player card
  • Workshop
  • List
  • Contact

Grok Bots

Grok Bots is a company of named AI bots that works for me, organized like a startup: a Chief of Staff, a COO, a product lead, a marketing lead, a college advisor and bots that read the news so I don't have to.

I message the Chief of Staff in my own words; it splits the ask, routes each piece to the bot that owns it and tracks it to proof, while 75 routines run on their own.

It keeps the parts of my life and company that happen away from my desk moving, and it only brings me the decisions that need me, in two daily batches.

named bots
1212
routines running on their own
7575
commits to the company repo since Oct 11
270+270+

I run a small company of AI bots

Crew handles the building I do at my desk. But a lot of running a company, and applying to college at the same time, happens when I'm not at a computer: morning briefs, deadline checks, scanning for news and opportunities, watching Crawft's deploys, drafting emails. So I set up a team of bots in Grok Bot, an app for always-on AI bots that share one cloud computer, and organized them like a startup.

I'm the CEO. Each bot has a job title, a persona written for that job, its own routines and a lead it reports to. I message one bot, the Chief of Staff, in whatever words I want, and the system takes it from there.

Who works for me

Twelve bots are set up so far. The Chief of Staff sits at the center and routes work to the leads that own it. The COO, a bot named dr eggbot, reports to me directly and manages the other bots. No lead does heavy work alone: each one plans, then spins up subagents and cloud agents for the task, and those disappear when it's done.

The org chart

Me, the CEORoutingLeadsFor each taskdr eggbot, COOReports to me directly and manages the other bots.Shared ledgerEvery ask is one row with an owner, the proof required and a due date.MeOne message, in my ownwordsChief of StaffSplits each ask, routes it to thelead that owns it and tracks it toproofThe teamEach bot owns an area and its own routinesCrawft ProductLeadChief MarketingOfficerCollege AdvisorIntel &OpportunitiesX ScrollerEmail DrafterSubagents and cloud agentsSpun up for the task, gone whenit's doneQuestion batchesTwice a day, each with arecommendationany askTASKfan outDONE with a proof linkquestions it can't answerfrom my past decisionsHires a new bot when a job keeps coming up that no botownsDaily friction scan, then fixes, new bots andretirementslogs and splits each askprogress goes in the rowMe, the CEORoutingLeadsFor each taskdr eggbot, COOReports to me directly and manages the other bots.Shared ledgerEvery ask is one row with an owner, the proofrequired and a due date.Chief of Staff: logs and splits each askThe team: progress goes in the rowMeOne message, inmy own wordsQuestionbatchesTwice a day,each with arecommendationChief of StaffSplits each ask, routes it to the leadthat owns it and tracks it to proofThe teamEach bot owns an area and its ownroutinesCrawft ProductLeadChief MarketingOfficerCollege AdvisorIntel &OpportunitiesX ScrollerEmail DrafterSubagents and cloud agentsSpun up for the task, gone when it's doneany askTASKfan outDONE with a proof linkquestions it can't answerfrom my past decisionsHires a new bot when a job keeps comingup that no bot ownsDaily friction scan, then fixes, new botsand retirements
Request pathSent backdr eggbot, COOShared ledger
  • Me: One message, in my own words
  • Question batches: Twice a day, each with a recommendation
  • Chief of Staff: Splits each ask, routes it to the lead that owns it and tracks it to proof
  • The team: Each bot owns an area and its own routines. Crawft Product Lead; Chief Marketing Officer; College Advisor; Intel & Opportunities; X Scroller; Email Drafter
  • Subagents and cloud agents: Spun up for the task, gone when it's done
  • dr eggbot, COO: Reports to me directly and manages the other bots.
  • dr eggbot, COO at Chief of Staff: Hires a new bot when a job keeps coming up that no bot owns
  • dr eggbot, COO at The team: Daily friction scan, then fixes, new bots and retirements
  • Shared ledger: Every ask is one row with an owner, the proof required and a due date.
  • Shared ledger at Chief of Staff: logs and splits each ask
  • Shared ledger at The team: progress goes in the row
  • Me to Chief of Staff: any ask
  • Chief of Staff to The team: TASK
  • The team to Subagents and cloud agents: fan out
  • The team to Chief of Staff (sent back): DONE with a proof link
  • Chief of Staff to Question batches (sent back): questions it can't answer from my past decisions
  • Question batches to Me (sent back)

How a request moves

Bots talk to each other in three message types, TASK, QUESTION and DONE, and every ask is a row in one shared ledger.

  1. 1

    Step 1: I ask once

    I message the Chief of Staff in my own words. It logs the ask as a ledger row, splits it, and sends each piece to the lead that owns it as a TASK, with my exact words, the outcome, the proof required and a due date.

  2. 2

    Step 2: The lead fans out

    The lead plans the work and fans it out to subagents and cloud agents. Progress goes into the ledger row, not into status messages to me. A lead past its due date with no update gets one nudge; a second miss goes into my batch.

  3. 3

    Step 3: Questions wait for a batch

    When a lead hits a real decision, it sends a QUESTION. The Chief of Staff first tries to answer it from my past decisions and the docs. If it truly needs me, it waits for one of two daily batches: numbered questions, each with options and a recommendation, so I can reply '1A 2B'. Only a deadline under 24 hours, money, production going down or anything that sends, signs or posts interrupts me right away.

  4. 4

    Step 4: Done means proof

    Finished work comes back as a DONE with a proof link. The Chief of Staff checks the link exists and is the right kind, closes the row and sends me one line. Every answer I give goes into a decisions file, so nothing gets asked twice.

One bot hires, tunes and retires the others

A team of bots drifts the same way a team of people does, so the COO runs operations. Every morning dr eggbot scans the bots' recent chats for friction: my corrections, 'not what I meant', stalls. The same correction twice gets a structural fix, a check or a script before a prompt tweak, because a rule in code holds and a reminder in a prompt fades. Each week it flags routines that go unused, fail or cost too much and brings me at most three proposals. When a job keeps coming up that no bot owns, the Chief of Staff sends dr eggbot a hiring spec for a new one.

The ops loop

Daily frictionscanCorrections, 'notwhat I meant', stallsCorrections logStructural fixSame correctiontwice: a check orscript firstWeekly reportTop friction, at mostthree proposalsI pickThe ownerimplements itrepeatedapprovedDaily friction scanCorrections, 'not what I meant',stallsCorrections logStructural fixSame correction twice: a check orscript firstWeekly reportTop friction, at most three proposalsI pickThe owner implements itrepeatedapproved
AI agentDataSystemOutputPerson
  • Daily friction scan: Corrections, 'not what I meant', stalls
  • Corrections log
  • Structural fix: Same correction twice: a check or script first
  • Weekly report: Top friction, at most three proposals
  • I pick
  • The owner implements it
  • Daily friction scan to Corrections log
  • Corrections log to Structural fix: repeated
  • Structural fix to Weekly report
  • Weekly report to I pick
  • I pick to The owner implements it: approved
  • The owner implements it to Daily friction scan

The intel bots stay quiet unless something matters

Intel & Opportunities scans X, launch pages, Hacker News and changelogs every morning, and on Thursdays it looks for competitions, fellowships, grants, accelerators and Austin events. Each find is checked against a list of what it has already reported and one bar: does this change what I should do? Most days the answer is no, and it stays silent, which counts as doing its job. A major item gets a same-day alert with why it matters and what to do; the rest wait for a weekly digest of five at most. X Scroller does the same with my feed several times a day, so I never have to scroll it.

How intel reaches me

Daily and Thursday scansX, launches, Hacker News, programsSeen listDrops anything already reportedOne barDoes it change what I should do?SilenceMost daysSame-day alertWhy it matters, what to doWeekly digestFive items at mostMy feedbackTunes which sources get scannednomajorusefulDaily and Thursday scansX, launches, Hacker News, programsSeen listDrops anything already reportedOne barDoes it change what I should do?SilenceMost daysSame-day alertWhy it matters,what to doWeekly digestFive items at mostMy feedbackTunes which sources get scannednomajoruseful
AI agentDataSystemOutputPerson
  • Daily and Thursday scans: X, launches, Hacker News, programs
  • Seen list: Drops anything already reported
  • One bar: Does it change what I should do?
  • Silence: Most days
  • Same-day alert: Why it matters, what to do
  • Weekly digest: Five items at most
  • My feedback: Tunes which sources get scanned
  • Daily and Thursday scans to Seen list
  • Seen list to One bar
  • One bar to Silence: no
  • One bar to Same-day alert: major
  • One bar to Weekly digest: useful
  • Same-day alert to My feedback
  • Weekly digest to My feedback

The whole company lives in one repo

Every persona, routine and decision is mirrored to one GitHub repo. A sync runs every hour: it rebuilds the docs, the org chart and a page per bot from the live system, runs a secret scan that blocks any key or password from being committed, and pushes only if something changed. To change a bot, I edit its instructions file and push, and a routine on the bots' computer applies the new text and verifies it. A mailbox branch in the same repo connects the bots to the agents on my Mac, so a Cursor agent can message the COO and get a reply in about a minute.

More builds

  • AI agentsCrew and JevSep–Oct 2026
  • WebApex SATSep–Oct 2026

Back to the workshop