IBM cut token usage 30x with agent logic — do we even need frontier models?
ibm's watsonx code assistant uses agent logic (static analysis/knowledge graphs) to reduce token usage by 30x while maintaining accuracy for legacy code tasks, suggesting scalable enterprise ai depends on software-defined guidance over raw model power.
ibm detailed how watsonx code assistant for z uses agent logic—software primitives like knowledge graphs and static analysis—to navigate complex enterprise workflows involving legacy cobol and pl/1 code. by pre-indexing application structure into database schemas, the agent retrieves precise information to guide the llm, resulting in roughly 30x lower token usage compared to a baseline frontier model while maintaining superior performance. this challenges the reliance on massive context windows and expensive frontier models, signaling that scalable enterprise adoption requires software-defined steering to manage dynamic workflows and regulatory constraints effectively.