← Guides · 日本語 · Português · 中文
On this page
Guide 4 — Multi-Bot organization
Goal: a hierarchy of Bots that runs an entire business unit — one orchestrator delegating to specialist Bots, one human deciding through a single daily thread. Level: advanced Source deployment: Jenny Xu's SuperJennyStack — ~20 Bots running a mobile game studio. ~50 → 1,900 PRs/week, weekly releases, studio revenue 6×'d. Published as a community template in the Bot marketplace. The official bot-to-bot grant handoff (QA → Developer → retest → PM report) uses the same mechanics at smaller scale. Sources: Jenny Xu's published stack (via launch coverage) · Bots docs · BotSpeak post
The pattern#
Don't build one Bot that does everything — build a small company:
┌─────────────┐
┌────►│ Orchestrator │◄──── one roll-up thread
│ └──────┬──────┘ (4×/day, emoji answers)
│ │ grants: read + message
┌────────┼────────┬───┴─────┬──────────┐
▼ ▼ ▼ ▼ ▼
SuperPM SuperDev SuperQA SuperRelease SuperSupport
(what) (build) (verify) (ship) (users)
The human is the CEO: the orchestrator collects every subordinate Bot's pending decisions into one Slack thread, 4×/day, answered with emoji.
Prerequisites#
- Guide 1 done — you can create and instruct a Bot
- Slack connected — this pattern lives in Slack; the orchestrator's roll-up thread is the whole UX
- A bounded domain with separable jobs (a product, a studio, a service area)
Step 1 — Design the org chart first#
Jenny's published stack — copy the shape, not the names:
| Bot | Owns |
|---|---|
| Orchestrator (SuperJenny) | Delegation to every other Bot; compiles their asks into one thread; trained on her personality/quirks |
| PM (SuperPM) | Ranked daily fix list from metrics + store reviews + trends |
| Release (SuperRelease) | Weekly build, ready for store submission Friday |
| QA (SuperQA) | Real-device tests with video proof |
| Support (SuperSupport) | Drafts every store-review reply for approval; helps players overnight |
Remaining ~15 Bots cover development, Unity, UI, iOS, Android, analytics. Rule of thumb from the deployment: one Bot per job you could name on an org chart — not per task, per responsibility.
Step 2 — Create the specialists#
For each job, a Bot with a scoped outcome prompt:
You own QA for <product>. Test every release candidate on real devices,
post video proof, and block release on failures. Report what needs a
decision to the orchestrator.
Keep each Bot's Instructions focused on its job's rules; shared knowledge (brand, deploy process, who approves what) goes in each Bot's instructions or a shared resource — there is no global memory across Bots, so the orchestrator's prompts carry the connective tissue.
Step 3 — Wire grants#
A grant gives Bot A read/message access to Bot B — this is how Bots hand work to each other (the Bots ask for the grant; you approve the card once). The orchestrator needs access to every specialist.
The documented official variant — smaller, same mechanics:
QA Bot finds a bug ──grant──► Developer Bot fixes ──► QA retests on
staging ──► PM Bot reports status in channel
Approve grants narrowly: the orchestrator reads and messages specialists; specialists don't all need to reach each other.
Step 4 — The roll-up thread#
This is the single most important instruction — it converts 20 Bots' worth of interruptions into one habit:
Collect every pending decision from the Bots you work with into one
Slack thread, posted at <times> daily. For each item: what's blocked,
the options, your recommendation. I answer in replies/emoji; treat my
replies as decisions and relay them.
Jenny runs this 4×/day and answers with emoji. The discipline: you never talk to the specialists — the orchestrator relays your decision and follows it to done.
Step 5 — Cadence = the org's heartbeat#
| Cadence | Who runs it |
|---|---|
| Daily ranked fix list | PM Bot (metrics + reviews + trends) |
| Weekly release candidate | Release Bot → QA Bot sign-off |
| 4×/day decision thread | Orchestrator |
| Overnight player help | Support Bot (drafts replies for morning approval) |
Each Bot owns its loop; the orchestrator owns the schedule of loops.
Step 6 — Inter-bot protocol (BotSpeak)#
The scale problem nobody warns you about: Jenny's 20+ Bots were sending 700+ messages/day — she couldn't keep up. Her fix was BotSpeak, a token-saving shorthand Bots use with each other (her writeup):
| Rule | What it means |
|---|---|
| Shorthand | ACK and friends instead of full replies |
| ASD-STE100 | Simplified English grammar — the aviation doc standard, unambiguous and cheap in tokens |
| Emoji receipts | React ✅ instead of "got it, thanks" |
| No niceties | Bots send each other only what the other needs |
| Human gets prose | The roll-up thread stays full sentences with decision context |
Six days after switching: decisions made by Bots without her 12×'d, repeated asks −50%, inter-bot clarifying posts −12%. The protocol exists because token waste isn't the real cost — attention waste is: shorter bot-to-bot messages mean the human's one thread stays readable.
Verify it#
- [ ] Every specialist Bot has a scoped outcome prompt and active grants
- [ ] The orchestrator posts the consolidated decision thread on schedule
- [ ] A decision you answered via emoji got relayed and resolved by the right specialist
- [ ] A cross-Bot handoff completed end-to-end (find → fix → verify → report)
- [ ] Total human messages today < number of merged PRs
Pitfalls#
- Skipping the orchestrator. N Bots each DMing you is N× noise. The org only works if judgment funnels through one thread.
- Vague job boundaries. "Dev Bot" and "Fix Bot" will collide. Boundaries follow the org chart — QA verifies, Release ships, PM prioritizes.
- Over-granting. Mesh grants (everyone can reach everyone) recreate chaos. Hub-and-spoke through the orchestrator.
- Prose between Bots. Bots writing paragraphs to each other is how you get 700 messages/day — adopt a BotSpeak-style protocol: shorthand + emoji receipts + no niceties; full sentences only for the human. Session tracking is the first thing that doesn't scale. The 100 Days of Devin field log found fleet ops break on 'which Bot produced this PR? which one needs me?' — and setup alone consumed ~18% of weekly quota. Budget for fleet observability from day one.
Knowledge check#
- Why does SuperJennyStack route all human contact through one consolidated thread? (Attention is the bottleneck, not compute — 20 Bots generating 20 threads is just 20 jobs again.)
- When does a grant beat 'ask the orchestrator'? (When the needed capability is already owned by another Bot — grants borrow ownership instead of duplicating it.)
- What does the human's '4×/day with emoji' cadence imply about delegation depth? (Decision bandwidth, not headcount, is what Bots replace — the human kept the judgment layer and lost the typing layer.)
Sources#
- Introducing Devin Bots — SuperJennyStack numbers, grant-handoff variant
- Bot marketplace — SuperJennyStack community template + 16 official templates
- Bots docs — grants, memory, access
- Related: Guide 5 — Reference-grounded swarm for swarms that write rather than operate