Multi-Agent: Direct Interaction

Deploy multiple agents, each owning a specific domain. You interact with each one directly through its own Telegram bot.

You are the orchestrator. The agents don’t talk to each other. Each one has clean, isolated context for its specific job.

The pattern

    You
   / | \  \
  v  v  v   v
[Photo    ] [Photo     ] [Print   ] [Print     ]
[Support  ] [Marketing ] [Orders  ] [Inventory ]
(Telegram)  (Telegram)   (Telegram) (Telegram)

Each agent:

• Has its own cage with dedicated secrets, storage, and configuration
• Connects to its own Telegram bot
• Knows only about its specific domain
• Hibernates independently — you only pay when it’s active

Example: Two-business setup

Running a photography studio and a screenprinting shop. Four agents, each with a clear responsibility:

Step 1: Deploy all cages

You’ll need a Builder plan ($49/month, up to 10 cages) or higher for 4+ agents. The Starter plan supports up to 3 cages.

lobster create photo-support
lobster create photo-marketing
lobster create print-orders
lobster create print-inventory

All four cages start in parallel. Each gets its own isolated environment.

Step 2: Configure each agent

Each agent gets its own AI key and platform token. More importantly, each gets a system prompt scoped to its domain.

# Photo support agent
lobster env set photo-support \
  ANTHROPIC_API_KEY=sk-ant-... \
  TELEGRAM_BOT_TOKEN=111111:AAA... \
  SYSTEM_PROMPT="You are the customer support agent for Bright Lens Photography Studio. You handle booking inquiries, session questions, print orders, and complaints. You know nothing about screenprinting — that's a different business."
 
# Photo marketing agent
lobster env set photo-marketing \
  ANTHROPIC_API_KEY=sk-ant-... \
  TELEGRAM_BOT_TOKEN=222222:BBB... \
  SYSTEM_PROMPT="You are the marketing manager for Bright Lens Photography Studio. You write social media posts, email newsletters, and ad copy. You know the brand voice: warm, professional, artistic. You know nothing about the screenprinting business."
 
# Print orders agent
lobster env set print-orders \
  ANTHROPIC_API_KEY=sk-ant-... \
  TELEGRAM_BOT_TOKEN=333333:CCC... \
  SYSTEM_PROMPT="You are the order management agent for Ink & Thread Screenprinting. You handle order status, shipping tracking, returns, and customer questions about their print jobs. You know nothing about photography."
 
# Print inventory agent
lobster env set print-inventory \
  ANTHROPIC_API_KEY=sk-ant-... \
  TELEGRAM_BOT_TOKEN=444444:DDD... \
  SYSTEM_PROMPT="You are the inventory manager for Ink & Thread Screenprinting. You track ink, blank garment, and supply levels. You alert when stock is low and coordinate reorders. You know nothing about the photography business."

The key detail: each system prompt explicitly excludes the other business. There’s no chance of context bleeding because the agent literally doesn’t have access to the wrong domain’s information.

Step 3: Set up each agent

For each cage, SSH in and run the interactive setup:

lobster ssh photo-support
openclaw configure
# Follow the prompts: choose AI provider, enter Telegram bot token
exit
 
lobster ssh photo-marketing
openclaw configure
# Same flow — each agent gets its own Telegram bot
exit
 
# Repeat for print-orders and print-inventory...

OpenClaw is pre-installed in every cage. The configure wizard handles AI provider setup, platform selection, and bot configuration. OpenClaw auto-starts on every wake from hibernation — no startup scripts needed.

Then get each cage’s webhook URL and register it with the corresponding platform:

lobster status photo-support    # → webhook URL for Telegram bot 1
lobster status photo-marketing  # → webhook URL for Telegram bot 2
lobster status print-orders     # → webhook URL for Telegram bot 3
lobster status print-inventory  # → webhook URL for Telegram bot 4

OpenClaw registers the webhook URL with Telegram automatically during openclaw configure. See the Telegram webhook guide for manual registration.

Step 4: Manage the fleet

See all your agents at a glance:

lobster ls

NAME              STATUS       SIZE      IDLE
photo-support     hibernated   starter   2h ago
photo-marketing   hibernated   starter   45m ago
print-orders      running      starter   active
print-inventory   hibernated   starter   3h ago

Check any individual agent:

lobster status photo-support -v

The cost reality

Four agents, each handling a few interactions per day:

That’s 1,500 credits out of your Builder plan’s 20,000 monthly allowance — under 8% of your included credits. Compare to four always-on VPS instances at $20-240/month.

The agents hibernate independently. If photo-support gets 10 messages in an hour and then nothing for the rest of the day, it’s active for ~15 minutes and hibernated for 23 hours 45 minutes. You pay for 15 minutes.

When to use this pattern

This pattern is ideal when:

• You have distinct domains that should never share context
• You want to interact with each agent directly through messaging
• Agents don’t need to coordinate with each other
• You want the simplest possible multi-agent setup

When to upgrade to the orchestrator

Consider the orchestrator pattern when:

• Agents need to pass data between each other (e.g., a research agent feeds results to a writing agent)
• You want automated task delegation without manual intervention
• You’re running many short-lived, task-specific agents
• You need programmatic control over the agent lifecycle