Language models that control software tools often call the model to make the same decision multiple times, wasting compute. HarnessMM stores each resolved action in a persistent graph. After a tool returns feedback, the system checks whether an already-stored decision can be reused and, if so, executes it without invoking the model again. The system also tunes its own configuration over many tasks.
On a standard code generation test, HarnessMM preserved the original success rate while using 29.54% fewer tokens. On a separate system, it cut token consumption by 87.93% while actually improving success. This shows that eliminating repeated reasoning can significantly lower the cost of running AI agents without compromising their performance.