A Telegram AI agent on its own is fine, but the real unlock is the master agent plus sub-agents architecture. Most Telegram AI agent tutorials show you how to build ONE agent and stop there. That's fine for casual users, but for real business use you need multiple agents working as a team.
This post is the architecture deep-dive that explains how to design and build the master plus sub-agents pattern that scales beyond hobby projects.
What A Telegram AI Agent Master Actually Is
A master agent sits at the top of your Telegram AI agent stack. It listens to every incoming message, decides which sub-agent should handle it, and routes accordingly.
You don't talk to the master directly (usually). It's the dispatcher that sits between your community and the specialists who do the actual work.
What Sub-Agents Are
Sub-agents are specialists. Each one has ONE job and they don't try to do anything else.
Examples include a Welcome sub-agent that greets new community members, a FAQ sub-agent that answers common questions, a Sales sub-agent that handles pricing and buying questions, a Support sub-agent that handles existing customer issues, a Spam Filter sub-agent that identifies and dismisses spam, and an Escalation sub-agent that routes complex issues to a human.
Each sub-agent has a focused system prompt and they don't try to do everything. That's exactly why they're good at what they do.
Why Architecture Matters
Single agents fail because they're trying to do too many things at once, their system prompts get bloated and contradictory, and they make decisions outside their area of expertise.
Master plus sub-agents fixes this in four ways. Each sub-agent has clear, narrow scope. The master agent's only job is routing. Each component is easier to test in isolation. You can update one sub-agent without breaking the others.
This is the same multi-agent pattern from OpenClaw computer use and ClawX OpenClaw, applied to Telegram.
Designing Your Telegram AI Agent Architecture In Three Steps
Three steps to design a working setup.
1 — List the message types you receive
Categorise the message types that hit your community. New member welcomes. Common questions (top 5). Pricing and sales questions. Existing customer issues. Spam and off-topic. Complex issues that need YOU.
2 — Map each type to a sub-agent
One sub-agent per type. Don't combine. Even if two types are similar, separate them so each sub-agent has clean scope.
3 — Build the master prompt
The master's job is simple: "For each incoming message, decide which sub-agent should handle it. Route accordingly."
Give it explicit rules. "If the message is from a new member, route to Welcome sub-agent." "If the message asks about pricing, route to Sales sub-agent." "If the message looks like spam, route to Spam Filter sub-agent." "If unsure, route to Escalation sub-agent."
🔥 Want my full master + sub-agent prompt templates? Inside the AI Profit Boardroom, I share my exact master prompts, sub-agent specs for community, sales, support, and escalation. Plus weekly live coaching where you can share your screen for help. 3,000+ members. → Get the templates
How Lobster Father Implements This
Lobster Father supports the master plus sub-agent pattern natively, which makes the setup straightforward.
The setup flow is to build each sub-agent as a separate agent in your platform (Tealaw, GPT agents, Lazy AI), create a master agent in Lobster Father, configure routing rules in the master, and let the master use your Lobster Father token to invoke sub-agents as needed.
The platform handles the orchestration. You handle the prompt design.
Four Common Architecture Mistakes
These four mistakes cause most multi-agent setups to fail.
1. Master doing real work. The master's only job is routing. If it starts answering questions itself, you've lost the architectural benefit.
2. Too many sub-agents. 5 to 7 sub-agents is plenty for most use cases. 20+ becomes hard to maintain and the routing logic gets brittle.
3. Overlapping sub-agent jobs. Each sub-agent should have ONE job. If two overlap, your master gets confused about routing and the whole system degrades.
4. No fallback. Always have a fallback (usually an Escalation sub-agent that loops in a human). If the master is uncertain, route to the fallback rather than guessing.
A Worked Example: My Community Manager Setup
Here's the real architecture I run on my Telegram community.
The master agent prompt is: "You're the dispatcher for Julian's Telegram community. For each message, decide: New member? → Welcome sub-agent. Pricing/buying question? → Sales sub-agent. Question about courses? → Course FAQ sub-agent. Existing customer issue? → Support sub-agent. Spam/off-topic? → Spam sub-agent. Anything else? → Escalation sub-agent (Julian)."
The Welcome sub-agent prompt is: "You greet new community members. Ask their name and what they're looking to learn. Suggest the most relevant intro resource. Keep it warm but brief."
The Sales sub-agent prompt is: "You handle pricing and buying questions. Refer to my pricing page. Don't quote numbers from memory. Direct serious buyers to book a discovery call."
The Course FAQ sub-agent prompt is: "You answer common questions about my AI Profit Boardroom courses. If unsure, escalate to Julian."
Each sub-agent is short, focused, and predictable. That's why the system works in production.
Scaling The Architecture
When your business grows, scale the architecture in three ways.
Add specialised sub-agents for affiliate questions, partnership requests, or podcast and media inquiries. Add tiered routing where the master goes to a Tier 1 dispatcher, then to a specific sub-agent, for high-volume operations. Add cross-platform agents that use the same master logic but route to agents on Slack, Discord, or email rather than just Telegram.
I cover the cross-platform side in Hermes Open Web UI.
Maintenance Cadence
Architectures need maintenance to stay sharp.
Weekly, read 10 random conversations, look for routing mistakes (master sent to wrong sub-agent), look for sub-agent mistakes (right sub-agent gave wrong answer), and update prompts to fix.
Monthly, review the routing rules, ask whether any sub-agents are underused, ask whether any sub-agents are overloaded, and adjust scope as needed.
How To Test The Architecture
Before going live, send 20 test messages of different types, confirm each routes to the right sub-agent, confirm each sub-agent gives a sensible response, and adjust where needed.
Don't deploy untested. The cost of a bad first impression on your community is way higher than the cost of a couple of hours of testing.
What This Architecture Doesn't Do
Be honest about the limits. It can't make complex business decisions. It can't replace nuanced sales conversations. It can't handle highly emotional customer situations.
For these, escalation to a human is the right answer.
When Single-Agent Is Enough
You don't always need multi-agent. Single agent is fine if you only have one clear use case (e.g. just FAQ), volume is low (under 10 messages a day), or you're testing the waters.
Once you're past that, master plus sub-agents is the upgrade path.
🚀 Want my full Telegram AI agent architecture templates? The AI Profit Boardroom has my exact master + sub-agent templates, daily training, weekly live coaching, and 3,000+ members. → Join here
FAQ — Telegram Master + Sub Agents
Do I need multiple sub-agents on day one?
No. Start with 2 to 3 and add more as you find gaps in coverage.
Can the master agent itself reply?
Yes, but it's a design smell. Keep master as router only.
How does the master decide where to route?
Based on rules in your master prompt. LLM-based routing usually works well in practice.
Will sub-agents see each other's conversations?
Depends on your platform. Most don't share by default, which is by design.
Can sub-agents call each other?
Yes. Telegram supports agent-to-agent communication on the new infrastructure.
How many sub-agents are too many?
Past 7 to 10 it gets hard to maintain reliably.
What if my routing is wrong?
Update the master prompt and iterate weekly. Routing improves with each pass.
Related Reading
- OpenClaw Computer Use — multi-agent desktop automation.
- ClawX OpenClaw — multi-agent OpenClaw setup.
- Hermes Open Web UI — Hermes channel management.
📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉
🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉
That's the architecture for a serious Telegram AI agent — master at the top, sub-agents specialised, escalation always available.
