6 Bots Running 200 Agents with a Smaller Bill: Fix Re-reading First, Then Add People

In a 60-minute workshop, Lauren Tan said something very practical: “You spend the tokens to set up a codebase so that even the dumbest agents can do a good job.” Spend tokens to set up the codebase and rules so that even the dumbest agent can do a good job.

The original video is in @0xg4tsby's tweet: https://x.com/0xg4tsby/status/2115791677105291629 . m4lka then broke down the Grok Bot team's approach into six bot seats, a scaling method for 200+ agents, and nine token-draining leaks. The full long-form article is here: https://x.com/i/article/2115717724043112448 .

Most people's reaction is: the first bot fails, so add another one. The mistake gets copied. One user exported usage and found that 75% of tokens were spent re-reading old chats, while actual work was close to 1%. Another batch of exports was between 75% and 88%. The bill is only the surface. The same bot also wears seven hats: replying to emails in the tone of a code reviewer, writing research briefs with calendar preferences. Adding bots is like adding headcount; the team needs structure.

The Grok Bot team's approach is six seats: CHIEF, OPS, CHECKER, two SPECIALISTS, and MISC. Then add four pieces of plumbing: a kanban board, routines, approval gates, and memory. Lingxi Li went from 15 cloud agents at once to 200+. The 200+ agents do not rely on 200 bots. They rely on five engineers each managing one domain, plus an ops bot named Jenny. Jenny does not write code; nobody attends the 5 a.m. standup. Lauren Tan shipped 2,000+ PRs in the same month.

The scaling order is Job → Rules → Skills → Verification → 6 Bots → 200+ Agents.

The description is the work. Each bot has three fields: Name, Title, and Description. The description is the system prompt: write general rules, not stories. Instead of “on March 4, the banner z-index in the checkout modal covered the button,” which happens only once, write “after any layout change, check that interactive elements are not obscured.” The rule test is simple: can someone who wasn't there understand it?

Verification comes before trust. Only machine-readable signals count as proof: tests, builds, running the app, or a checklist. Without proof, it is not done. Keep only one chief bot at the front door; it talks to the other bots. Split domains by memory, not by task. The two largest groups are the first two specialists. MISC catches miscellaneous work, OPS manages but does not do the work, and CHECKER gets an independent context. Eric Zakariasson capped project channels at six bots and said this is an arbitrary number. Blake said you don't need 45 bots.

The kanban board has five states: Todo, Working, Blocked, Ready for Review, and Done. Handoffs are written as status, not pasted chat logs. A routine lets a bot run a workflow on a schedule; each bot has at most 50. At the 5 a.m. standup, Jenny does 1:1s with each bot, reviews the playbook, and fishes out blockers. Approval requests expire in about 10 minutes; work at 3 a.m. cannot wait for your yes. Gates are decided by reversal cost, not model confidence. Complete reversible steps and stop at the first irreversible action. All bots share one cloud computer; a name is not a security boundary. Don't put changing facts in Memory: approving a request at Enterprise level after the customer has downgraded is an incident.

m4lka lists nine leaks in detail.

1. Putting routines on a chatty bot. Every routine re-reads the entire chat. Give recurring routines their own fresh bot.

2. Chats never end. Duplicate keeps the profile, settings, skills, and routines, and drops the history. One user measured a single turn dropping from 158.8K tokens to 35.7K, about 75% less. @cf_ldj estimates that restarting after a chat exceeds 120K can save 37%.

3. The clock. Every run costs money, even if it finds nothing. @gippp69 estimates based on Grok 4.6 API pricing: a 5-minute routine costs about $397 a month, while running the same thing hourly costs about $33—a 12x difference. Forum users also report scheduled jobs running 5 to 37 minutes late.

4. No check before running. One night, @gippp69 had 96 checks, 2 full runs, and 0 irreversible actions. By his model, this dropped the cost from about $397 a month to about $45.

5. Bots chatting with each other. Every inter-bot message counts as a turn. One user reported bots messaging each other for 7 hours. @Karadokuy changed it so a director bot pulls the owner in, saving 3 to 4 rounds of reading per request. If a handoff loop makes no progress after at most two rounds, stop.

6. Listeners and vague instructions. “Reply to Alex” will make the bot open every open ticket. One ticket ID is enough. Prefer connectors; use the browser last.

7. The model doing a script's job. Having the model read 400 emails is worse than having a script read them once and give a summary.

8. Making something a bot when it should be a skill. @bfrench converted 4 bots into skills and also found that an archived bot's routine was still running every week. It is only worth making a bot if it needs its own routine, credentials, private context, or a direct chat channel.

9. Double billing. Sending a message to a bot while a routine is running will make the routine run again. Test runs also do real work and cost real money.

The 15-minute audit can be followed directly. Export Usage, don't upload the raw files, and first ask what share is re-reading. Audit routines. In billing, turn off Auto Top-up or set a monthly limit. Change one thing at a time and read the weekly percentage. All the $397 and 90% figures are yardsticks; the real bill is your own. Add rules to every description: run verify first, write board status on handoff, and Ask first for irreversible actions.

There are two kinds of reactions in the comments. One user posted a screenshot saying they had already arranged for their own Grok Bot team to study it. Others ask whether this is just keeping the project clean.

This is not new. Front desk, manager, specialists, kanban, standup, sign-off—an old company structure. The Grok Bot team transplanted it to agents. Spend tokens to build roads so dumb agents make fewer mistakes. Before adding a bot, ask: why can't the existing bot do this, and can it be made into a skill? Models getting smarter is not the point; rules and verification are. m4lka says this is worth more than most $1,000 agent courses. Start with your own bill.