How My 7 AI Agents Run Board Meetings (And How to Build Your Own)

Last Updated: March 5, 2026

A multi-agent AI system uses specialized AI agents with distinct roles, reasoning techniques, and data access to collaboratively make business decisions. Instead of one AI chatbot that forgets you every conversation, seven agents — CEO, CFO, CMO, CTO, COO, Research, and Critic — run structured board meetings inside a Telegram group, each contributing independent analysis before a human reviews and approves final decisions. This is the Autonomous Entrepreneur project, built on Claude Code and running 24/7 on a self-hosted server. More than 366+ professionals in the Autonomee community are already building their own versions.

I want to show you exactly how it works. The architecture. The Telegram setup. The reasoning techniques behind each agent. And how you can start building your own.

The Problem: 10 AI Tools, Zero Business Intelligence

You probably have ChatGPT for writing. Notion AI for tasks. Perplexity for research. Claude for coding. A calendar tool. A meeting summarizer.

Five, six, ten tools. None of them talk to each other.

You spend $200 or more a month on subscriptions. And every time you open a new chat, you start from zero. No memory. No context. No visibility into your business.

I am not even talking about your CRM data, your revenue numbers, your content pipeline. All that rich business intelligence sitting in silos.

According to Gartner’s 2026 AI Agent Report, 40% of enterprise applications will feature task-specific AI agents by end of 2026, up from less than 5% in 2025. The shift from scattered tools to centralized AI systems is not a trend. It is a structural change in how businesses operate.

The approach of fixing this with one centralized system is what we call AI Second Brain. It knows you. It knows your business. And it takes action.

We call ours GoBot.

What GoBot Actually Does

GoBot is my secure, self-hosted replica of OpenClaw. It reads my emails, checks my calendar, knows my revenue, my code base, my content pipeline, projects, tasks. It runs 24/7 on my own server. I have full control of the infrastructure. The data stays mine.

But the feature I use every single day — the one I am most proud of — is board meeting.

Seven specialized agents collaborate in a structured board meeting process. Each one brings a different lens, a different reasoning technique, and access to different business data. One of them exists solely to break everyone else’s arguments.

Meet the 7 AI Board Meeting Agents

Each agent has a specific role, a specific reasoning technique, and access to specific tools and data. Here is what each one does and how it thinks.

1. The Orchestrator (Chief of Staff)

The general agent. It handles everything that does not need a specialist. Think of it like your personal assistant coordinating with different board directors. It is adaptive, flexible, and routes complex questions to the right expert. During board meetings, it runs the synthesis phase — reading all agent responses and extracting themes, conflicts, and action items.

2. Research Agent — ReAct Reasoning

ReAct stands for Reason then Act. This agent does not just generate an answer. It searches, fact-checks, and quotes the source. Every claim comes with a reference. It uses WebSearch and WebFetch tools for market intel, competitor analysis, and trend research. The reasoning cycle goes: Plan, Search, Analyze, Repeat — until it has verifiable answers.

3. Content Agent (CMO) — Recursion of Thought

My CMO thinks about packaging, audience psychology, and trends. If I ask it about a video idea, it does not say “great idea Goda, you are a genius.”

It looks at everything I have ever done, what is performing well, and tells me which angle would work and why. It uses Recursion of Thought reasoning — iteratively drafting, critiquing, and refining its own output before presenting the final recommendation.

4. Finance Agent (CFO) — Chain of Thought

Chain-of-thought math. Step by step. ROI calculations, unit economics. Very conservative with money. If a decision costs too much, it says so. It has full access to my actual revenue, transactions, and financial accounts.

5. Strategy Agent (CEO) — Tree of Thought

This is my second favorite. Tree of Thought reasoning means it does not give you one answer. It explores multiple futures, multiple paths, and picks the one with the highest upside. It thinks like a venture capitalist. Contrarian by design. Projects 3-month, 1-year, and 3-year futures.

6. CTO — Technical Architecture

Handles infrastructure. Checks my GitHub. Opens pull requests, tracks issues, helps develop projects. If something breaks in the code base, CTO knows before I do.

7. COO — Operations

Tasks, deadlines, calendar, SOPs, compliance. The operations backbone that keeps everything running and everyone accountable.

The Critic: My Favorite Agent (And Why It Matters Most)

The Critic is a devil’s advocate. Its entire job is to question things and break things. If Strategy says “Hey Goda, your community is growing, raise prices” — the Critic says “What if 50 people leave because of that?”

It cannot be optimistic. I baked in my Eastern European no-bullshit approach. No hype. Critical thinking only.

If you know about the AI benchmarks where we test models on nonsense questions — Sonnet 4.6 and Claude models generally do not comply with your nonsense. They call it out. Now imagine combining that built-in honesty with an additional brutally honest critic persona on top.

Full superpowers. This is the agent that prevents groupthink and catches blind spots the other six miss.

Why Most Multi-Agent Systems Are Broken (The Silo Problem)

Most multi-agent AI orchestration frameworks have a main orchestrator on top that controls multiple agents underneath. But if those agents cannot communicate with each other directly, you have the exact same problem you have in business. Silos.

According to Microsoft’s research on designing multi-agent intelligence, “multi-agent systems can be modeled as graphs, with agents represented as nodes” — but the architecture of those connections determines whether agents collaborate or just report up (Source: Microsoft Developer Blog, 2026).

In GoBot, each agent can communicate with each other and invoke one another. Not just report to the top. This is the architectural decision that separates a real board of advisors from seven separate chatbots that happen to be in the same app.

How Agents Invoke Each Other

Any agent can invoke another agent mid-conversation using a simple tag:

[INVOKE:finance|What's the break-even on this strategy given current MRR?]

The system parses the tag, loads the target agent’s full context, gets a response, and posts it in the same Telegram thread from the target agent’s bot.

Three safeguards prevent infinite loops:

1. An agent cannot invoke itself
2. Maximum invocation depth of 2 levels
3. Chain tracking prevents circular calls (strategy to finance to strategy is blocked)

The Telegram Architecture: How to Wire It Up

All seven agents live in a Telegram supergroup with Forum Mode enabled. Each agent has its own Telegram bot token and its own forum topic. The architecture is simple and visual:

Telegram Supergroup (Forum Mode)
├── General     → Orchestrator bot
├── Research    → Research bot
├── Content     → Content bot
├── Finance     → Finance bot
├── Strategy    → Strategy bot
├── CTO         → CTO bot
├── COO         → COO bot
└── Critic      → Critic bot

When I send a message in the Finance topic, it routes to the Finance Agent’s system prompt, reasoning technique, and tools. When the finance agent responds, it comes from the finance bot. When the critic responds, it comes from the critic bot. Visually distinct. Contextually isolated.

Step-by-Step Telegram Setup

Here is how to set up the multi-bot system:

1. Create bots in BotFather — one per agent (e.g., @YourResearchBot, @YourFinanceBot, @YourCriticBot)
2. Create a Telegram supergroup with Forum Topics enabled in group settings
3. Create one topic per agent — Research, Content, Finance, Strategy, CTO, COO, Critic
4. Add all bots to the group as admins so they can post in any topic
5. Configure environment variables for each bot token:

TELEGRAM_BOT_TOKEN=your_main_bot_token
TELEGRAM_BOT_TOKEN_RESEARCH=research_bot_token
TELEGRAM_BOT_TOKEN_CONTENT=content_bot_token
TELEGRAM_BOT_TOKEN_FINANCE=finance_bot_token
TELEGRAM_BOT_TOKEN_STRATEGY=strategy_bot_token
TELEGRAM_BOT_TOKEN_CTO=cto_bot_token
TELEGRAM_BOT_TOKEN_COO=coo_bot_token
TELEGRAM_BOT_TOKEN_CRITIC=critic_bot_token

6. Map topic IDs to agents in your routing config:

const topicAgentMap: Record<number, string> = {
  3: "research",
  4: "content",
  5: "finance",
  6: "strategy",
  7: "cto",
  8: "coo",
  9: "critic",
};

When a message arrives in a topic, the system looks up the corresponding agent, loads its system prompt, reasoning technique, and tool permissions, and responds from that agent’s bot.

The BotRegistry manages all bot instances, handles Telegram’s 4096 character limit by chunking at paragraph boundaries, and falls back to the primary bot if a specialized token is not configured. So you can start with one bot and add more as you go.

How AI Board Meetings Actually Work (The 6-Phase Process)

I can trigger a board meeting by typing /board or naturally saying “I want a board meeting about launching a new product.”

Phase 1: Context Gathering

The system pulls context from the past 7 days — 30 messages per topic. It retrieves previous board meeting decisions (only confirmed, human-approved ones). It resets the session state so each meeting starts clean.

Phase 2: Agent Contributions (Sequential, Independent)

Each agent presents independently, in order: Research, Content, Finance, Strategy, CTO, COO, Critic. Each receives its own system prompt, board context, previous decisions, and live data relevant to their role. A typing indicator shows in Telegram while each agent “thinks.”

I do not want cross-pollination during this phase. Each agent gives its perspective based on their own data and expertise. Independent analysis first.

Phase 3: Critic Analysis

The Critic is the only agent that sees all other agents’ full responses. It asks: if everyone agrees, what are we missing? Why are we agreeing? What is another perspective we have not considered? Same approach as in a real boardroom with real people.

Phase 4: Synthesis

The Orchestrator reads all responses and synthesizes themes, conflicts, alignments. It proposes 3-5 action items with clear ownership.

Phase 5: Human in the Loop

This is my part. The board meeting stays open for me to review, ask questions, provide clarifications, correct the agents. I can go back and forth with all of them in one group. If I do not respond within 30 minutes, the meeting auto-closes. Smart safeguard against forgotten sessions.

Phase 6: Decision Extraction

When the meeting closes, the system extracts only final outcomes: decisions (what was agreed, with my corrections overriding agent proposals), corrections (what agents got wrong that I corrected), and action items (next steps with ownership).

The Two-Layer Memory System (Why This Works Across Sessions)

This is the architecture detail that makes multi-agent board meetings sustainable over time.

Layer 1 (Raw Transcript): Full conversation stored in a Supabase database. Every agent’s input, every discussion point, every correction. This is for reference only.

Layer 2 (Decisions Only): Only human-approved outcomes. With explicit corrections of what agents got wrong.

Future board meetings only read Layer 2. Not the full discussion. Not raw agent proposals.

Why does this matter? If you store raw proposals as “decisions,” future meetings re-propose the same ideas. Agents hallucinate consensus that never happened. When you close a real board meeting with real people, you do not bring the entire 2-hour discussion to the next meeting. You bring decisions and open items. Same principle. Clean signal. No noise.

What Community Members Are Building

Dan in the Autonomee community took the board meeting concept and applied it to creative projects. His “Creative Impact Board” has agents focused on virality, longevity, value, and scalability. They debate with controlled temperature settings. A final decision-maker runs the output through what he calls a “vibe matrix” — scoring ideas as green light, yellow light, or red light.

Other members have integrated this architecture with WhatsApp, Slack, and Discord. The system is platform-agnostic. Telegram is where we started, but the agent logic and board meeting process work on any messaging platform that supports bot APIs.

Some members use local models for sensitive financial data. Others run everything on VPS. The architecture supports full self-hosting — your server, your data, your rules.

Privacy and Self-Hosting Options

If your CFO agent has access to real revenue data (and it should), you need to think about where that data flows.

Options the system supports:

• Self-host on VPS — your server, your data, full control
• Self-host on your machine — no external server at all
• Local models for memory — use Ollama or similar for storing and retrieving sensitive context locally
• API only for reasoning — send anonymized context to Claude for the thinking, keep raw data local
• Hybrid model routing — use Haiku for simple queries ($0.003/msg), Sonnet for medium tasks, Opus for deep analysis

Full autonomy. Full independence. Your infrastructure, your choice.

Frequently Asked Questions

How much does it cost to run a multi-agent AI board meeting system?

The base infrastructure runs on a VPS for approximately $5-10 per month. API costs for Claude depend on usage — board meetings with seven agents typically cost $0.50-2.00 per session depending on complexity and model tier. The system supports smart model routing: Haiku for simple queries, Sonnet for medium tasks, Opus for deep analysis. Most users spend $30-60 per month on API costs total, replacing $200+ in scattered tool subscriptions.

Can AI agents actually make good business decisions?

The agents do not make decisions. They analyze, present perspectives, and challenge assumptions. The human in the loop — you — makes the final call. This is the same structure as a real board of directors: advisors present, the founder decides. According to research from Unite.AI, business leaders increasingly view AI agents as “tools to empower employees, not as autonomous decision-makers that operate in isolation” (Source: Unite.AI, 2026).

Do I need to know how to code to build this?

The free Claude Telegram Relay requires basic terminal commands — cloning a repo, editing environment variables, running a server. No programming needed for the basic setup. For the full multi-agent system with board meetings, the Autonomee community provides step-by-step courses and live help from 366+ members who have already built theirs.

How do you prevent AI agents from hallucinating in board meetings?

Three mechanisms. First, each agent contributes independently during the board meeting — no cross-pollination that could compound errors. Second, the Critic agent specifically challenges consensus and flags unsupported claims. Third, the two-layer memory system only stores human-approved decisions, not raw agent proposals. Future meetings read confirmed decisions, not hallucinated consensus.

Can I use this with tools other than Telegram?

The agent logic and board meeting process are platform-agnostic. Community members run the same architecture on WhatsApp, Slack, Discord, and custom web interfaces. Telegram is the default because of its excellent bot API, forum topics for visual agent separation, and supergroup capabilities. But the core system — agent definitions, reasoning techniques, cross-invocation, memory layers — works on any platform that supports bot messaging.

How to Get Started

If you have a clear vision for how AI agents should work in your business, the hardest part is not the code. It is understanding your own use case. Where is your data? What decisions do you make repeatedly? Where are you the bottleneck?

Your job is to understand your business. AI does the heavy lifting.

The free Claude Telegram Relay is MIT licensed and always will be — yours forever.

For the full GoBot with all 25+ features including board meetings, multi-agent orchestration, voice calls, proactive AI, and everything I showed here — that is what we build together inside Autonomee.

366+ professionals are already building their AI Second Brain. 97.9% stick around month after month. That tells you something.

Join us inside Autonomee.