Dex Horthy presents 12-Factor Agents as a practical framework for building LLM agents that are reliable in production. His main argument is that good agents are not magic. They are software. That means developers should directly own the critical parts of the system rather than outsourcing them blindly to frameworks.
1. Agents Fail When Abstractions Hide Too Much
Many teams can get an agent to 70 or 80 percent quickly, but then reliability collapses. At that stage, debugging becomes painful because prompts, tools, state, and execution paths are hidden inside opaque abstractions.
Horthy argues that the answer is not more magic, but more ownership.
2. Core Principles
The summary highlights several key ideas:
- prompts and context windows should be explicitly managed,
- tool use is just structured input and output, not a mystical capability,
- control flow should be owned in deterministic code,
- execution state and business state should be separated,
- and human interaction should be designed deliberately into the loop.
He also favors micro-agents: small LLM loops embedded inside larger deterministic systems.
Conclusion
The framework's spirit is simple: reliable agents come from the same kind of engineering discipline that makes any software reliable. If you own the prompts, state, context, and control flow, you can build systems that are easier to understand, debug, and trust.
