Editor's note: “Lobster” refers to an OpenClaw agent in the original article. The configurations, account outcomes and cost figures below are examples reported by the author and named users; they have not been independently reproduced. The source's monthly-cost example totals $135 and reports $65 savings, implying a $200 comparison and 32.5% savings, while a later quotation claims 60%. Both statements are retained as in the original.
Original · Unique Research · 2026-02-24
COVER STORY
Multi-Lobster Collaboration in Practice: From Three Lobsters Fighting to an AI Army
I had 3 little lobsters run one project simultaneously, and they really started fighting.
Last week, my friend Aaron placed three lobsters in the same Discord server: one for the frontend, one for the backend, and one for documentation, using Forum channels to keep sessions separate. Half an hour later, the backend lobster deleted a component the frontend lobster had just written; the documentation lobster @-mentioned him in the group and said, "the requirements changed"; and the frontend lobster threw out a pile of incomprehensible error logs. The three lobsters each acted independently, and the project immediately collapsed into chaos.
This is not a joke, but a snapshot of what I found after compiling experiences from 29 real users.
When you evolve from "raising one lobster" to "raising a group of lobsters," the real bottleneck is never "how to start a few more," but "how to stop them from fighting among themselves."
I distilled these painful lessons into 5 of the most commonly used collaboration methods. Some people used them to raise Reddit account-warming success from 30% to 80%; others used 200 lobsters to maintain social-account matrices for clients. One of these approaches should suit you.
First, the most basic and effective approach: separate "experimentation" completely from "real work."
Bruce is a typical example. He raised one lobster called "Steady" on a Mac mini and another called "Try It" on a Tencent Cloud VPS in Silicon Valley. The two lobsters did not know each other at all. "Try It" crashed ten times, while "Steady" continued sending its daily report on schedule every day.
How did he do it?
Step One: Two Independent openclaw.json Configurations
"Steady's" production environment loads only 3 stable skills: daily-report, feishu, and calendar. autoUpdate is shut off completely, and it never upgrades automatically. "Try It's" test environment uses a skill allowlist of ["\u002a"] (the JSON escape for the asterisk wildcard), with experimentalFeatures fully enabled and sandboxMode turned on.
Step Two: Physically Isolated Data Directories
Mac mini: ~/.openclaw-stable/
VPS: ~/.openclaw-sandbox/
The two directories never access each other. Bruce even configured different Telegram Bot Token credentials for them to prevent messages from crossing between systems.
"I once tested a new crawler skill and it crashed the browser—but the daily-report task in my production environment was not affected at all." As Bruce said this, his tone carried the relief understood only by someone who has stepped into that pit.
Next, many people encounter a second problem: one person must simultaneously serve "personal life" and "company business."
Bella's solution was simply to raise two lobsters with completely different personalities.
The one in Vegas was called "Little Zhang," with soul.md stating, "You are a freelance developer who loves experimenting with new technology." The one in the Bay Area was called "Mr. Zhang," stating, "You are the technical lead; stable delivery comes first."
The key: two completely independent workspaces.
workspace/
├── personal/ # "Little Zhang"
│ ├── soul.md # freelance-developer persona
│ ├── user.md # personal-project background
│ └── openclaw.json # aggressive configuration: Claude Opus model, temperature 0.9
│
└── company/ # "Mr. Zhang"
├── soul.md # technical-lead persona
├── user.md # company-business background
└── openclaw.json # conservative configuration: GPT-4 model, temperature 0.3
Bella even connected the two lobsters to different Feishu groups: "Little Zhang" appeared only in the "personal projects group," and "Mr. Zhang" only in the "company business group."
"The same model and the same underlying capabilities produce work with completely different characters because the personas and parameters differ," she said.
Building on this, some people took the approach to an extreme. Julian raised three personality lobsters at once:
"Rigorous Lobster": soul.md emphasizes "coding standards, unit tests, and edge conditions."
"Creative Lobster": soul.md states, "You excel at brainstorming and are not constrained by convention."
"PM Lobster": soul.md states, "You excel at breaking down tasks and pay attention to delivery milestones."
He then gives the same requirement to all three and has each propose a solution. Finally, he manually combines them—often producing results far better than a single lobster.
At the next level, some people begin using "channel isolation" to avoid risk controls.
Victor wanted to build a Reddit matrix. One central-control lobster handled brain-level decisions, with Docker lobsters and a physical iPhone beneath it.
His architecture looked like this:
Central-Control Lobster, or Hub
Deployment: local Mac mini
Responsibilities: receive Feishu instructions, split tasks, and aggregate results
Configuration: maxSpawnDepth=5, allowing derived subtasks up to 5 levels deep
Docker Lobsters 1 & 2, the PC Executors
Deployment: cloud VPS with Docker isolation
Independent IP: each container binds to a different proxy, Proxy1 or Proxy2
Independent browser fingerprint: Puppeteer in stealth mode
Physical iPhone, the Mobile-Side Executor
Deployment: a real iPhone, not a simulator
Purpose: simulate real users' mobile behavior
Victor's central-control lobster received an instruction in the Feishu group: "Post 3 items today." It then invoked the fanOut skill and split the task into:
Docker Lobster 1: post 1 technical item, sub-session 1
Docker Lobster 2: post 1 lifestyle item + interact, sub-session 2
iPhone: browse + like, sub-session 3
The three sub-sessions executed in parallel and sent aggregated results back to central control after completion.
"Previously, one lobster managed every account, and 3 were banned in half a month. It has now been two months without a single one dying, and the activity data look especially natural." Victor's data: account-warming success rose from 30% to 80%, while account survival increased from 7 days to more than 60 days.
Another group prioritizes computing power and pursues extreme cost-performance.
Ivan directly prepared a separate machine for compute-heavy projects. Max was smarter: a hybrid architecture.
Max's cost calculation:
Configuration | Monthly Cost | Purpose | AWS g5.xlarge (GPU)
$120
Video generation and large-scale data analysis, running every day before dawn
Mac mini M4
$15 in electricity
Replying to messages, checking weather, and simple search, responding at any time
Total $135; saves $65 versus all-cloud while adding unlimited computing power versus all-local
Max's key design was task routing:
He defined separate capability boundaries for the "cloud lobster" and "local lobster" in soul.md:
Cloud Lobster, or CloudBot
Good at: video generation, data analysis, and scheduled tasks
Constraint: nonurgent tasks that may be delayed
Trigger words: "generate video," "analyze data," and "schedule"
Local Lobster, or LocalBot
Good at: immediate replies, simple queries, and calendar management
Constraint: lightweight tasks with responses in seconds
Trigger words: "now," "immediately," and "check"
When a user message arrives, Max first evaluates the keywords, then decides which lobster should handle it.
"Personal matters go local; company tasks go to the cloud. It saves 60% in monthly costs and responds faster as well," Max said.
The final approach is a "role-collaboration matrix" that truly treats AI as a team.
This is also the most advanced approach currently. Aaron ultimately transformed the three lobsters into a genuine "development team."
His Discord server structure:
#general ← Coordinator Lobster, Coordinator
#frontend ← Frontend Lobster, Frontend
#backend ← Backend Lobster, Backend
#docs ← Documentation Lobster, Docs
Key configuration for the coordinator lobster:
{
"label": "coordinator",
"role": "coordination",
"canDelegate": true,
"maxSpawnDepth": 3,
"skills": ["task-splitter", "progress-tracker"]
}
Configuration for the frontend/backend/documentation lobsters:
{
"label": "frontend",
"role": "execution",
"canDelegate": false,
"parentLabel": "coordinator",
"skills": ["react-coder", "ui-designer"]
}
Key mechanism: parallel Sub-session work
For the requirement "build a login page," the coordinator lobster will:
1. Open sub-session 1 in #frontend: "write the React login component"
2. Open sub-session 2 in #backend: "write the login API endpoint"
3. Open sub-session 3 in #docs: "write the API documentation"
The three sub-sessions execute in parallel while the coordinator lobster monitors progress with the progress-tracker skill. When sub-sessions 1 and 2 are both marked "complete," sub-session 4 is triggered automatically: "integrate the frontend and backend code."
Aaron showed me their collaboration records. Every day, the coordinator lobster sends a "project daily report" listing the progress, blockers, and next-step plan for every subtask.
"These three lobsters can now deliver a small project in full. The coordinator breaks down tasks and monitors progress, the frontend and backend lobsters develop in parallel, and the documentation lobster writes the API documentation in sync—it is smoother than coordinating it myself." As Aaron said this, the three lobsters were running a new feature in the background. We chatted in the group while they quietly did the work.
From Chaos to Order, Separated Only by "Relationship Design"
With one lobster, you are raising a tool; with a group of lobsters, you are raising a team.
A tool only needs to work. A team must know who answers to whom, how data flows, and how conflicts are resolved.
The worst case I saw involved Xavier's two lobsters issuing different instructions for the same task. A new employee received two contradictory onboarding guides, while the same task in the IT system was marked both "completed" and "pending." The reason was that both lobsters monitored the same Feishu group, without a primary-subordinate division.
I also saw BTS's 4 lobsters reply to a user simultaneously in the same group, causing the user to ask in frustration, "Which one of you is actually talking to me?" BTS later gave each lobster a different @ trigger: @PA_Scheduler, @Mentor_Advisor, @PM_Planner... Only then was the conflict resolved.
But I also saw the strongest example: HaLoHa uses 200 lobsters in orderly fashion to maintain clients' social accounts, each lobster handling 3-5 accounts under unified central scheduling for scaled delivery. Victor used channel isolation to increase Reddit account-warming success from 30% to 80%.
The difference lies in whether one question was considered clearly at the beginning: what relationships exist among these lobsters?
Want to Get Started? Begin with the Simplest Step
Step One: Identify the Separation Point
Look at your usage scenario. Which separation point hurts most?
Personal vs work? → Follow Bella: two workspace+soul.md systems
Stable vs experimental? → Follow Bruce: two openclaw.json+ physical-isolation configurations
Light tasks vs heavy computing? → Follow Max: cloud-local hybrid + task routing
Step Two: Start a Second Lobster to Do Only This One Thing
Do not begin with 5 lobsters. First implement the clearest separation point:
Keep the existing configuration locally, the primary lobster
Start a new cloud instance running only scheduled tasks, the secondary lobster
Step Three: Design Collaboration Rules
If two lobsters need to collaborate, clarify three questions:
1. Which is primary and which secondary? The primary lobster can schedule work; the secondary only executes.
2. How does data flow? Shared directories vs message passing.
3. How are conflicts resolved? Different trigger words vs different channels.
Step Four: Expand Gradually
After two-lobster collaboration works smoothly, add a third and fourth lobster as needed, along with more complex "relationships."
Closing Thoughts
At this point in writing, I suddenly realized: the ultimate objective of raising lobsters is never "the more, the better," but "the more they collaborate, the stronger they become."
As in managing a team, 10 people acting independently may be less efficient than 3 people collaborating seamlessly. Lobsters are the same.
Before deploying a second lobster, consider its relationship to the first: how will they collaborate?
Only after thinking this through can your AI army truly function.
How many lobsters are you raising now? What is the most outrageous episode of "infighting" you have encountered?