The abstract for TreeFlex, posted on arXiv, identifies a core problem with ReAct-style agents: they append every observation, intermediate result, and failed attempt into a single context window. Over time, that context becomes bloated with stale or irrelevant details, which competes for limited capacity and can obscure the evidence the agent actually needs.
The proposed solution, implied by the title and framing, is to have agents navigate their own search trees rather than a flat history. By structuring past actions and findings as a tree, the agent can more selectively revisit useful branches and avoid revisiting dead ends. The abstract stops short of providing experimental details, but the motivation is clear: context hygiene matters for long-horizon reasoning.
This is a single-source report, so there are no conflicting findings to compare. The paper appears to target a known weakness in current agent designs, though the abstract alone does not show how much improvement the method yields in practice.