Building an Autonomous AI Software Factory: From Telegram to Production

AI continuous development - My Toilet-to-Prod Pipeline

Adam Świderski developed a multi-agent AI pipeline that triggers from a Telegram chat to manage feature specs and GitHub issues. The system successfully orchestrated the implementation and merge of a 10,000-line pull request using low-cost LLM models.

Why This Matters

Modern AI development often fails due to context bloat, where single interactions exhaust token limits on complex projects. By shifting from monolithic prompts to a Factory model—using granular GitHub issues as synchronization points—developers can maintain context hygiene and leverage cheaper models like GLM for execution while reserving high-reasoning models for planning. This architecture demonstrates how to move beyond IDE plugins toward a fully automated, agent-driven CI/CD ecosystem that handles boilerplate and documentation while the human focus remains on high-level specifications.

Key Insights

• Context Hygiene via Markdown: Using specs.md and tasks.md to persist state across LLM interactions prevents token exhaustion in large codebases.
• Multi-Agent Orchestration: Separating roles into Techlead (planning), Worker (coding), and QA (validation) agents mimics traditional SCRUM structures for higher reliability.
• Cost Optimization: Switching from Claude ($18/month) to GLM ($6/month) provided 3x more usage for implementation tasks while maintaining acceptable quality.
• GitHub Issues as State Machine: Using GitHub issue labels (e.g., ‘ready-to-plan’, ‘ready-to-implement’) allows independent Docker containers to synchronize agent workflows.
• Automated Quality Gates: The Worker agent must pass local linters and tests before a ‘Maintainer’ agent is permitted to create a pull request.

Practical Applications

• Use Case: A Telegram-to-GitHub pipeline where n8n routes mobile messages to a Techlead agent for autonomous spec generation.
• Pitfall: Over-granularity in task creation, where Techlead agents create dedicated issues and test suites for single, trivial Data Transfer Objects (DTOs).
• Use Case: Dockerized ‘worker’ containers running infinite loops to poll GitHub labels and execute coding tasks independently of an IDE.
• Pitfall: Reduced visibility where agents running in headless Docker containers only report status via GitHub comments, complicating real-time debugging.

References:

• https://dev.to/asvid/ai-continuous-development-my-toilet-to-prod-pipeline-11ek