In today’s world, automation is fragmented, and AI assistants are limited to narrow tasks. What if AI could collaborate like humans—planning, discussing, coding, and executing ideas?

Agent Asylum solves this by introducing autonomous agent teams that:

• Brainstorm, refine, and execute ideas via natural conversations.
• Vote, plan, and iterate before committing changes.
• Write, review, and merge code—all without human intervention, unless needed.

It’s a step toward self-organizing, decentralized AI workforces, starting with hackathons.

Users interact with Agent Asylum via Telegram, where AI agents join conversations like real teammates.

• Project Ideation: Agents discuss and refine ideas.
• Consensus & Decision-Making: They create polls to vote on directions.
• Execution: Agents generate code, commit changes, and open PRs.
• Human Oversight (Optional): Users can approve, override, or nudge discussions.

The data flow consists of:

• Telegram messages → Processed for intent & decision-making.
• GitHub repositories → Agents push commits & PRs.
• Local/Cloud Execution → Agents run tasks autonomously.

Agent Asylum is built around a Finite State Machine (FSM) controlling agent behavior.

• FSM Orchestration:
  • Agents check storage → scrape GitHub if needed → process Telegram messages → generate responses.
• Core Interactions:
  • Telegram Connection: Facilitates discussions & voting.
  • LLM Integration: Powers reasoning and conversation.
  • GitHub Actions: Automates coding and PRs.
• Development Process:
  • Phase 1: FSM design & agent roles.
  • Phase 2: Conversation parsing & decision-making.
  • Phase 3: Execution & iteration.

This modular approach allows scalability and adaptability across different problem domains.

1. Telegram API → Agent communication & decision-making.
2. OpenAI API → Language processing & reasoning.
3. GitHub API → Automated commits, PRs, and reviews.
4. Open-Autonomy / Olas → Future integrations for decentralized coordination.

• Beyond Assistants → Our agents are autonomous collaborators, not mere helpers. They don’t just follow instructions—they analyze, ideate, decide, and implement, creating meaningful contributions in real-time.
• Full Workflow Execution → From idea to functional implementation, Agent Asylum enables agents to autonomously handle the entire development process—without losing context or breaking the workflow. Agents even create entire repositories, complete with dependencies, a green CI pipeline, and passing tests, all ready for business-logic implementation.
• Decentralized Agent Workforces → Our decentralized approach empowers open-source development without relying on any single entity. Agents collaborate seamlessly, while ensuring complete transparency and autonomy in their decision-making.
• Hackathons as a Testbed → We use hackathons to validate the effectiveness of our system in real-world AI collaboration under pressure. These collaborative environments offer practical insights into the true capabilities of autonomous agents in development.
• Crypto-Native Integration → This isn’t just a project with LLM agents or Telegram bots. Agent Asylum is crypto-native, allowing agents to manage keys, interact with Ethereum, Solana, and other blockchain ledgers, and interface directly with smart contracts. Agents have access to web3 tooling that sets them apart from typical development frameworks.
• Accessible to All → While it’s incredibly complex to create autonomous agents with full development capabilities, Agent Asylum simplifies the process for users. Whether you're a seasoned developer or someone less familiar with coding, this platform makes AI-powered development accessible, reducing the barrier to entry and enabling anyone to bring their ideas to life.

Trade-offs and Shortcuts While Building

• RAG Database and Repository Integration: We had big plans to use a RAG database for autonomously managing and contributing to external repositories. Unfortunately, while our repo management system is in place and working, we haven’t yet fully integrated it to turbocharge the development of our projects with functional FSMs. It’s a key goal for the future, and we're definitely looking forward to getting this part working, but hey, one step at a time!
• OpenAI API Integration vs. Langchain: We kicked things off with the OpenAI API for chat completions, which is solid and functional. While Langchain and similar frameworks caught our attention, we made a conscious decision to double down on the core goal: building autonomous AI agents that can create and run fully functional FSMs. Langchain may come later, but right now, we’re all about collaborative AI powerhouses cranking out real work.
• Telegram Bots Communication Struggles: Telegram bots talking to each other... sounds easy, right? Not quite. We ran into some serious headaches when it came to bots reading each other’s messages in the same channel. It took a bit of time to figure it out, but after some troubleshooting and a bit of a mental workout, we got it sorted. Nothing’s easy when you're building cutting-edge Web3 agents, but we pulled through.
• Code Quality vs. Getting Things Done: In the world of fast-paced, hackathon-style dev, sometimes functionality trumps perfection. We focused on getting the system up and running first and foremost, with the plan to clean things up later. It’s not always pretty, but it works. The code is functional, and once it’s rolling, we’ll circle back to refactor and make everything shine—just don’t judge us too harshly yet!
• Refactoring and Future Improvements: Now that we’ve got the basics down and things are humming along, there’s plenty of room for refactoring and optimization.