Tool Synthesis

Pain Signals — You Need This Pattern When:

• The same class of LLM task runs hundreds or thousands of times with the same logic
• Results should be deterministic but LLM outputs vary across identical inputs
• Inference cost is accumulating for tasks that do not require reasoning
• Latency is unacceptable for a task that should complete in milliseconds
• An LLM is doing data transformation, validation, formatting, or calculation — tasks code handles better
• The LLM is a bottleneck in a pipeline where most steps are deterministic

Core Principle

An LLM call is interpretation. A generated script is compilation. Tool Synthesis is the LLM recognising that a task has become well-understood enough to compile into a deterministic artifact — then writing that artifact and stepping out of the loop.

The LLM does not simply generate code. It makes a deliberate architectural decision: should this task remain probabilistic reasoning, or should it be codified as a reusable tool? The decision itself requires judgment; the resulting tool does not.