The Four Levels of Agent Maturity

Building AI agents is often plagued by "mass psychosis"—the confusion of whether to use complex frameworks or manual coding. The author proposes a four-level maturity model to navigate this:

  1. Framework Prototyping: Use existing frameworks (e.g., LangChain, LangGraph) only for initial Proof of Concept (PMF) testing. These are useful for rapid validation but lack the modularity and customizability required for production-grade systems.
  2. State-Machine Architecture: Move to custom-built agents. Treat every agent as a recursive state machine—a while loop with specific conditions and end states. Visualizing the agent as a state machine makes it easier to debug and manage complex logic.
  3. Kanban UX for Inference-Bound Agents: Because agents are often inference-bound (running for 8–10 minutes), the ideal interface is a Kanban board. This allows a human to act as an "engineering manager," overseeing multiple parallel agents that mutate source code in isolated environments.
  4. Cloud-Native Fleets: For scaling, move agents to the cloud. This removes local dependencies, enables parallel execution of long-running tasks (e.g., 60-minute Q&A cycles), and allows for shared, consistent environments across teams.

Five Rules for Writing Agent Code

To avoid building "slop," follow these five technical heuristics:

  • Prune System Prompts: Frontier models are increasingly capable; excessive instructions lead to sensory overload and degraded performance. Keep prompts minimal.
  • Think in State Machines: Always maintain a clear mental model of the agent's current state within its recursive loop.
  • Build for AI-to-AI Interaction: Design your agent’s CLI and CI/CD pipelines so that other coding agents can easily test, modify, and deploy your agent. If it is hard for a human to test, it is hard for an AI to improve.
  • Architect Manually: While AI can write code, the high-level architecture and state-machine design must be handled by a human to ensure the system is thoughtful and coherent.
  • Respect API Specifics: Frontier labs often lock down APIs. Failing to use exact formats for features like "reasoning traces" results in silent performance degradation. Do not rely on generic wrappers like OpenRouter if you need peak performance.

The Future of Agent Interaction

As agents become more autonomous, the human role shifts from direct guidance to management. By using Kanban boards, users can track agent progress, handle state transitions (e.g., moving a task from "In Progress" to "Review"), and manage multiple parallel threads. Moving these workflows to the cloud allows for "fire and forget" execution, where an agent can perform complex tasks—like building a VS Code extension, configuring settings, and running terminal tests—entirely in the background.