Story

I’ve spent the last couple of years deploying LLM agents in production environments, and I’ve consistently hit the same wall: the 'Cognitive Monolith' (or what I call the Big Ball of Mud AI).

We are currently seeing a lot of hype around 'autonomous agents' that are essentially 3,000-word prompts with access to 20 tools. In my experience, this doesn't scale. It’s impossible to unit test, observability is a nightmare, and the 'vibe-based' engineering makes it a liability for enterprise-grade software.

I wrote this manifesto to formalize a different approach: Microagentic Stacking (MAS).

The core idea is to apply classic software engineering principles—separation of concerns, strict I/O contracts, and atomic modularity—to the agentic stack. Instead of one god-agent, we build a stack of 'micro-agents' that:

1. Have a single, specialized responsibility. 2. Communicate through typed, validated interfaces. 3. Are individually testable and replaceable.

I’m sharing this as an open-source manifesto because I believe we need to move from 'prompt alchemy' to 'agentic engineering' if we want these systems to be more than just cool demos.

I’d love to hear the community’s thoughts on: - How you handle state management across multiple specialized agents. - Where you see the trade-off between modularity and token latency. - If you’ve found better ways to prevent 'agentic sprawl' in complex workflows.

The repo includes the core principles and a conceptual roadmap. Happy to dive into the technical details.