Production-tested templates for building AI agent teams on OpenClaw with Telegram supergroup integration.

This is an OpenClaw-specific project. These templates are designed to run on the OpenClaw platform — they use OpenClaw's agent system, workspace structure, session management, sessions_send for bot-to-bot communication, and Telegram channel bindings. If you're not running OpenClaw, these won't work out of the box.

Built from a real production setup running 10 autonomous agents coordinated through a Telegram supergroup with dedicated topic channels. Each agent has its own bot, personality, workspace, and domain expertise.

Agent personality patterns and the concept of specialized AI agent divisions are inspired by The Agency by @msitarzewski — a collection of 55+ specialized AI agent personalities built for Claude Code. Great resource for agent design philosophy, personality crafting, and workflow patterns regardless of what platform you run.

What we build here is different: the operational infrastructure for running those kinds of agents as a live, coordinated team on OpenClaw — with Telegram supergroups, shared context files, bot-to-bot triggers, cron schedules, topic-based team channels, and structured escalation chains.

A complete template kit for deploying a multi-agent team on OpenClaw + Telegram:

• SOUL.md templates — Agent personality and behavior definitions
• IDENTITY.md templates — Agent metadata and capabilities
• Workspace templates — Shared context, memory, and coordination files
• openclaw.json snippets — Configuration for agents, bindings, channels, and teams
• AI Instructions — Step-by-step setup guide written for AI agents to follow
• Architecture docs — How agents communicate, escalate, and share context

                    ┌──────────┐
                    │  You 👑   │
                    └────┬─────┘
                         │
                    ┌────▼─────┐
                    │Lead Agent│ Orchestrator
                    └────┬─────┘
          ┌──────────────┼──────────────┐
          │              │              │
    ┌─────▼─────┐  ┌────▼────┐  ┌─────▼─────┐
    │ RESEARCH  │  │  BUILD  │  │  MARKET   │
    │  Team     │  │  Team   │  │  Team     │
    └─────┬─────┘  └────┬────┘  └─────┬─────┘
          │              │              │
    Research +     Code → QA →    Content +
    Analytics      Deploy          Community

• Three Telegram routing models — multi-bot routing, native topic routing, and DM forum topics
• Multi-bot routing — Each agent has its own Telegram bot token and visible identity
• Native topic routing — One Telegram bot can route different topics to different internal agents via topics.<id>.agentId
• One topic per team — Teams share a topic channel in a supergroup
• Primary + Secondary agents — Primary owns the topic; secondary responds when mentioned or triggered
• Shared context via files — Agents coordinate through shared markdown files, not APIs
• Structured escalation — Clear rules for when to escalate up the chain

In native topic routing, agentId controls the internal OpenClaw agent that handles the message — workspace, memory, tools, prompt, model, session.

It does not control the visible Telegram bot identity.

So this:

"13": { agentId: "connor" }

means Topic 13 is handled by the connor agent internally, but replies still come from whichever Telegram account is attached to that group/topic (often the default/orchestrator bot).

Use this rule of thumb:

• Want one visible bot with many internal specialist brains? Use native topic routing.
• Want Connor to look like Connor and Kara to look like Kara? Use multi-bot routing.

See docs/supergroup-setup.md for the full step-by-step guide.

See INSTRUCTIONS.md — a complete setup guide written specifically for AI agents to follow. Give it to your orchestrator agent and let it handle the setup.

See templates/soul/README.md for the full role catalog, adaptation checklist, and advanced OpenClaw practices.

OpenClaw v2026.3.24+ includes a native Skills system with one-click install from ClawHub, Control UI management, and CLI tools. All templates above are compatible with the native system. See docs/skills-system.md for details.

Topic: Single shared topic Agents: Coder (primary) + QA (secondary) + DevOps (secondary) Flow: Coder builds > triggers QA via sessions_send > QA tests > triggers DevOps > DevOps deploys

Topic: Single shared topic Agents: Researcher (primary) + Growth Analyst (secondary) Flow: Researcher posts findings > triggers Analyst for metrics/experiments

Topic: Single shared topic Agents: Content Creator (primary) + Community Manager (secondary) Flow: Content posted on Twitter <> adapted for Reddit, cross-pollination of insights

Telegram bots cannot see each other's messages. This is a Telegram limitation. Use OpenClaw's sessions_send to trigger teammates.

Use the enforced handoff format documented here:

• Inter-Agent Handoff Standard

Quick example:

sessions_send(agentId="qa-agent", message="HANDOFF\nfrom: coder\nto: qa\ntask_id: build-142\npriority: P1\nsummary: Validate checkout fix\ncontext: branch=fix/coupon-rounding\ndeliver_to: telegram:-1001234567890:13\ndeadline: asap\ndone_when:\n- Repro no longer fails\n- Regression checks pass")

The receiving agent gets the message as a new session input and posts ACK + DONE updates in the shared topic.

Agents can delegate coding work to dedicated coding agents (Claude Code, Codex, OpenCode) via ACPX:

1. Define a coder agent with runtime.type: "acp" in openclaw.json
2. Other agents trigger it via sessions_spawn(runtime="acp") from a subagent session
3. Or use /acp spawn codex --bind here to turn any Telegram conversation into a coding workspace

See docs/acpx-telegram.md for the full guide.

• examples/full-team.json — Complete 10-agent team config
• examples/minimal-team.json — Minimal 3-agent setup (orchestrator + coder + QA)

• INSTRUCTIONS.md — AI-readable setup guide (give this to your agent)
• Agent Design Patterns — How to write effective SOUL.md files
• Supergroup Setup Guide — Step-by-step Telegram setup
• Telegram Channel Architecture — Production lane design for topic routing
• Telegram DM Topics Guide — How to create and use forum topics inside a direct chat
• Scaling Your Team — Adding new agents and teams
• Advanced OpenClaw Practices — High-leverage patterns for real multi-agent ops
• Skills System — Native OpenClaw skills, ClawHub marketplace, creating custom skills
• ACPX for Telegram — Using ACPX as a Telegram coding agent backend

PRs welcome. If you build a useful agent template or team configuration, share it.

MIT

Agent personality patterns inspired by The Agency by @msitarzewski. OpenClaw operational layer by @raulvidis.