The experiment
We tested Ramain on realistic long-horizon tasks. With no prior knowledge, the agent completed every task, but slowly and with detours. The baseline was about 6.3 minutes and 44 model rounds per task.
Then Ramain learned from its successful trajectories, in two rounds.
Round one: extract the workflow
Workflow extraction turns successful trajectories into procedural guidance. Wall time fell 68%, from 6.3 minutes to 2 minutes per task. Fresh input tokens dropped by nearly half.
Round two: turn the procedure into muscle memory
An optimization agent wrote reusable scripts from the trajectories. These subroutines do routine steps in code, with no model round at all. Total model rounds fell from 885 to 438. Against the baseline, tasks ran 79% faster and used one third of the fresh input tokens.
Production performance compounds
The first deployed version of an agent is its least experienced one. Each later run draws on more procedural knowledge, so the agent gets faster and cheaper as the workflow repeats.