Long-horizon language agents keep accumulating reasoning history, and every old chain of thought is re-sent with each new request. This work proposes ICLR, a training-free online method that scores reasoning blocks with a frozen proxy entropy model and compresses low-scoring blocks while keeping actions, tool calls, and observations intact. On 260 WorkBuddyBench tasks, mean reward rose from 0.699 to 0.718, and input, output, and cache read tokens fell by 25.5%, 14.4%, and 33.3%.
The key caveat is that agent reasoning is not static text. Deleting one reasoning block can change the next action and therefore the whole trajectory, an effect the authors call trajectory amplification. Additional analyses with representation probing and activation patching suggest reasoning matters less once the relevant derived state has been written to code, files, tool outputs, or environmental feedback.
The result matters because it reframes reasoning in agents as dynamic working state, not permanent history. It suggests context management for long-horizon agents can be done cheaply, without retraining, and that externalizing state is what makes old reasoning safe to forget.