Repeated attempts at interactive tasks generate facts, useful operations, and failure cases that can guide later executions. We introduce the ReGE framework, which turns recorded experience into guidance for a fresh attempt. Its main implementation, ReGE‑Syn, requires only a single coach call to jointly synthesize bounded execution records and available artifacts into self‑contained guidance, linking task‑specific findings with proposed actions and potential pitfalls. A student agent starts from the task's initial environment state and adapts this guidance to current observations. Source generation, synthesis, and student execution all use the same base model, without parameter updates or a stronger teacher. Across four benchmarks, ReGE improves accuracy by about 9.9 % for GPT and 8.8 % for Claude compared with unguided execution. On matched BrowseComp‑Plus subsets, ReGE‑Syn reduces guidance‑construction tokens by 98.2 % (GPT) and 95.5 % (Claude) relative to an adapted AggAgent, while maintaining success. Combined construction and execution tokens (counting input and output, excluding source generation) drop by 15.7 % and 35.4 %, respectively. Ablation studies on BrowseComp‑Plus show that access to process information leads to fewer student steps, and joint synthesis yields guidance that is substantially shorter than Direct history or Independent summaries, with comparable success. These findings demonstrate that agents can leverage their own recorded experience to improve subsequent task completion and suggest a low‑cost way to consolidate execution records of parallel agent systems such as agent swarms.
Review