My LLM kept calling tools it shouldn't, so I built a state machine to stop it

Dev.to / 6/11/2026

💬 OpinionDeveloper Stack & InfrastructureIdeas & Deep AnalysisTools & Practical Usage

共有:

Key Points

The article describes a real failure case where an LLM confirmed a booking even though the database showed it was already sold.
It explains that prompt-based safeguards are insufficient because the underlying issue is that restricted tools can be present in the LLM’s context during the wrong conversation step.
To prevent incorrect tool use, the author built a declarative state machine that controls which tools are available per conversation state.
The proposed approach, implemented in the zero-dependency TypeScript library reactive-fsm, ensures the model cannot call tools unless the state machine has enabled them for that phase.

Continue reading this article on the original site.