GrokBot Loops

Source

Recirc: Codez (@0xCodez) — x.com/0xCodez/status/2095185610390573536 (2 Sep 2026). ~55 min video. Overlay: “99% of people using GrokBot just for 1% of its real power. They run 1 agent without loop & graph.” Quotes his 18 Aug article (same article as /grokbotelon) . Earlier Codez cut (27 Aug) named the stack as /loop, /goal, /swarm inside pstack.

No new transcript. Same Maven/Lenny room as /grokbotteam (Lauren × Colin, water-clap, Glass, P-Stack, Dune). Full ASR lives there. This page is the study cut: what “loop” means, and one you can run tonight.

Replies on the recirc: unlimited tokens; 20 agents → 100% usage in three minutes. Pair with /grokbotlimits.

Video (the recirc file)

Open the tweet / watch on X → · 462p amplify. Prefer /grokbotteam if you want the session with a searchable transcript.

One-sentence TL;DR

A loop is a job that runs, checks a real stop rule, and comes back. A graph is several of those jobs with named owners and handoffs. Lauren’s /loop is Cursor P-Stack hill-climbing a verification skill — not a Grok Bot button labeled Loop.

Three things people call “loop”

Overlay “without loop & graph” is Codez packing (2) and (3) into one scare line. Lauren’s actual ladder is: verification skill → eval → /loop the eval → cloud agents → then many bots. Skip the first rungs and you get 20 untitled chats.

Her words, from the ASR on /grokbotteam:

An eval is a unit test for an agent. … you can hill climb them, meaning that your eval can produce a score. … you can use slash loop in Cursor and you can say, okay, keep looping on this eval until everything is 10 out of 10. … a judge agent of a different model [to] make sure that the first model is not being biased.

That’s Cursor. On Grok Bot you fake the same shape with two bots and a routine — paste block below.

What she actually built (from the talk we already transcribed)

1. Trust curve: you cannot parallelize what you don’t trust.
2. Verification skill (Control Glass + feature map) so the agent runs the product.
3. Howl: stop guessing, read the code.
4. Evals as unit tests for skills. /loop until the rubric is green. Judge model ≠ worker model.
5. Local first, then Cursor cloud, then Benny (bug reports while she sleeps).
6. Dune architecture: shortest path is the right path; CI bans the footguns; auto-merge only after that.

/goal and /swarm are P-Stack playbooks on the coding side (Codez’s 27 Aug overlay). They are not Grok Bot sidebar items. On Grok Bot the equivalent is: one goal in the description, a routine as the clock, a second bot as the swarm-checker.

How to learn it this sitting

Don’t watch 55 minutes of overlay twice. Do this:

1. 20 min — read /grokbotteam (trust curve through Dune). Skip the water-clap ASR.
2. 10 min — map the three loops on this page. Write one sentence: “My loop’s stop rule is ___.” If you can’t name a checkable stop, you don’t have a loop. You have a cron.
3. 30 min — one tiny Grok Bot loop, read-only. Worker: “open this URL / this folder, write a 10-line status file to /workspace/loop-status.md.” Watcher (or Chief): “if the file exists and the timestamp is newer than last run, ping me; else stay quiet.” Schedule hourly, not every 5 minutes.
4. Only then point the same shape at a real job (Ad Library Scout, GSC page-2, bookmark digest). Write actions stay behind Require Approval.

Graph comes last: two bots in a group chat with one owner per stage. That’s CasJam Head + Reviewer, not 20 workers.

Paste this tonight (read-only)

Two bots. One folder in /workspace. Hourly clock. Nothing sends, spends, or pauses ads.

Worker description

You are Loop Worker. Job: once per run, inspect the assigned source (URL or folder). Write /workspace/loops/status.md with: time (ISO), what you checked, what changed vs the previous file if it exists, and one next action. If nothing changed, say so in one line. Do not send email, post, or change any live system. If you are unsure, say unsure — do not guess.

Watcher / Chief prompt (after the first manual run)

You are Loop Watcher. You do not do the Worker’s job. Read /workspace/loops/status.md. If the timestamp is newer than the last note in /workspace/loops/log.md, append one line to log.md: time, changed yes/no, and whether the Worker followed the description. Ping me only if something changed or the Worker skipped the file. Stay quiet otherwise. Never mark the work “done” because the Worker said it was done — the file is the stop rule.

Run Worker by hand once. Correct the file format. Then: schedule Worker hourly; Watcher after Worker, or as a second routine on the hour + 10 minutes. If you only have one bot tonight, Watcher can be Chief. Don’t add a third bot.

A loop that isn’t a furnace

Related on this site

• GrokBot team — the Lauren talk + full ASR (start here)
• Graph engineering — jobs, arrows, shared state (Greg Isenberg)
• SEO loop — Access / Action / Memory / Judgment / Reversal
• Field manual — ladder, maker-checker, quota furnace
• CEO · org — Chief + channels, not 20 untitled workers
• Limits — 20 agents will not fit a normal week
• Riley — MTS every 5 min; steal the list not the clock
• Stop babysitting agents · TOC

Primary recirc: @0xCodez — Lauren / loop & graph · speaker @poteto · P-Stack cursor.com/marketplace/cursor/pstack