We have audited dozens of AI agent projects that died in the first month. The pattern is always the same.
## It is not the model
Teams blame the LLM. "Claude is not smart enough." "GPT-4 hallucinated." The model is rarely the problem. The problem is what we asked it to do.
## It is the documentation
Most agents are built on top of a process that was never written down. The human builder has it in their head. The agent does not.
So when an edge case appears, the agent guesses. When a tool call fails, the agent gives up. When the input format changes, the agent breaks.
## The fix
Write the documentation first. Trigger, inputs, decision points, edge cases, outputs. In builder-ready language. Then build the agent against the documentation, not against your memory.
Agents built this way last for years. Agents built without documentation last for weeks.