Search for a command to run...
Series
Experiments in moving work out of the LLM and into deterministic code, using a tennis court booking agent. Each post compares agent designs on the same scripted conversations and reports pass rate, tokens, cost and latency. Post 1 compares a ReAct agent, a prompt-graph agent and a simple workflow API agent. Post 2 replaces tool calling with Jev as the turn interpreter.
The most reliable architecture is the one that asks least of the model. Summary Conversational agents are often built by writing a workflow into the system prompt and letting the model work out which
Replacing the tool-calling LLM with a jev classifier cut total cost from $0.2866 to $0.0265 and p90 turn latency from 6,421ms to 3,508ms on the same 11 scripted conversations. Both agents passed 11 of
