Multi-Agent Learning — Three Progressive Projects

A learning journey through multi-agent LLM orchestration, building progressively from a CrewAI sequential pipeline → a hierarchical CrewAI crew → a hand-rolled Anthropic ReAct loop deployed as an interactive Telegram bot.

All three projects use Claude Haiku 4.5 as the underlying LLM.

🤖 Try the live Telegram bot: @TruthComitteeBot

Send it a controversial claim or a numeric question. The Chief Editor will dispatch the Investigator, hand the file to the Analyst, crunch numbers with the Calculator, and deliver a verdict — narrating each step. Running 24×7 on a Google Cloud free-tier VM.

What's in here

File	What it is	Stack
main.py	Tech News Pipeline — a 2-agent sequential CrewAI crew (Researcher → Writer) that drafts a blog post about Anthropic's Claude Managed Agents.	CrewAI + LangChain (DuckDuckGo search)
truth_committee.py	The Truth Committee (CrewAI) — a 3-agent hierarchical crew (Lead Investigator + Skeptical Analyst, managed by a Chief Editor) that produces a balanced Risk-Benefit Report on humanoid robotics in manufacturing by 2027, with confidence scores per claim.	CrewAI hierarchical process
TruthComitteeBot.py	The Truth Committee (Telegram bot) — same concept, rebuilt without a framework. Uses the Anthropic SDK directly with a hand-written ReAct loop. Multi-user, async, streams status updates as the committee deliberates.	python-telegram-bot v20+, AsyncAnthropic, ddgs

The progression

1. Sequential (main.py) — Researcher hands off to Writer in a fixed order. Linear pipeline. Easiest mental model.
2. Hierarchical (truth_committee.py) — A Chief Editor decides who works next and can loop back for revisions when the Analyst flags gaps.
3. Native ReAct, no framework (TruthComitteeBot.py) — Drop CrewAI entirely. Hand-written async ReAct loop with explicit tool dispatch, per-role iteration caps, programmatic backstops, and Telegram streaming for many concurrent users.

The three files together demonstrate why you might use a framework like CrewAI early on (fast to a prototype) and why you might eventually drop it (token bloat, opacity, less control over iteration logic).

Local setup

git clone <this-repo-url>
cd <repo-folder>

python3 -m venv .venv
.venv/bin/pip install --upgrade pip
.venv/bin/pip install -r requirements.txt

cp .env.example .env
# Edit .env — paste your real ANTHROPIC_API_KEY (and TELEGRAM_TOKEN if running the bot)

Get keys:

• Anthropic API key: https://console.anthropic.com/settings/keys
• Telegram bot token: message @BotFather on Telegram, /newbot, follow prompts.

Running each project

.venv/bin/python main.py             # Tech News Pipeline (CrewAI sequential)
.venv/bin/python truth_committee.py  # Hierarchical Truth Committee (CrewAI)
.venv/bin/python TruthComitteeBot.py # Telegram bot (native Anthropic ReAct)

main.py and truth_committee.py run once and print to stdout. TruthComitteeBot.py is a long-running poller — Ctrl+C to stop.

Deploying the bot to Google Cloud free tier

The Telegram bot is configured to deploy on a Google Cloud e2-micro VM (Always Free in us-central1, us-east1, or us-west1). See deploy/README.md for step-by-step instructions, including the systemd unit that keeps the bot running 24×7 across reboots.

Architecture notes for TruthComitteeBot.py

The interesting bits, condensed:

• No framework. ~600 lines including comments. Direct AsyncAnthropic + python-telegram-bot.
• ReAct loop — while iter < MAX_EDITOR_ITERATIONS: call Anthropic with the Editor's system prompt and tools; on stop_reason == "tool_use", dispatch each tool, append results, repeat; on end_turn, send the final reply.
• Per-role caps (ROLE_CAPS) — at most 3 investigator calls, 3 analyst calls, 3 calculator calls per user message. Once both investigator and analyst caps are hit, the next API call drops the tools list, forcing the model to synthesize a verdict.
• Per-chat asyncio.Lock + history dict — many concurrent users, but a single user's rapid double-message can't race their own history. Different chats run in parallel.
• Status-first ordering — the user sees 🕵️ Dispatching Investigator… before the slow tool call returns, so the chat feels alive.
• Safe calculator — pure ast-based, never calls eval(). Whitelisted nodes only (no Names, Calls, Attributes).
• Markdown → Telegram-HTML converter so Claude's **bold** and ## headers actually render in Telegram.
• Concurrency gotcha — python-telegram-bot v20+ processes updates serially by default. To get true parallelism across users, the Application is built with concurrent_updates(True).

File structure

.
├── main.py                       # Project 1: CrewAI sequential
├── truth_committee.py            # Project 2: CrewAI hierarchical
├── TruthComitteeBot.py           # Project 3: native Anthropic + Telegram
├── requirements.txt              # All deps for all three projects
├── .env.example                  # Template — copy to .env and fill in
├── deploy/
│   ├── README.md                 # GCP deployment walkthrough
│   ├── setup.sh                  # One-shot VM setup script
│   └── truth-committee-bot.service  # systemd unit
└── README.md                     # this file

License

MIT — do whatever you like with it.