代理評測
GDPevo Tests Agent Self-Evolution Through Rule Recombination, but Full Prompts Still Lead Autonomous Learning by More Than 30 Points
GDPevo decomposes enterprise workflows into atomic rules and recombines them in held-out test tasks to determine whether agents can accumulate transferable strategies from previous tasks. Public results show that autonomous evolution does deliver gains, but it remains far below the 91.6% ceiling achieved with access to the complete rules.
Agent “self-evolution” is often implemented through accumulated memory, skill files, or workflow modifications. But higher test scores do not necessarily mean that an agent has genuinely learned new rules; it may simply be encountering similar tasks again. GDPevo, released on August 4, attempts to narrow this attribution problem. The research team decomposed workflows spanning CRM, ERP, finance, healthcare, legal services, data analysis, and engineering operations into atomic business rules. It distributed subsets of those rules across five training tasks, then recombined them into five held-out test tasks. The test tasks therefore do not directly repeat the training tasks, but still require agents to transfer process knowledge acquired earlier.
The repository currently includes 24 sets comprising 240 tasks, along with the data-construction pipeline, a rule-based evaluator, and reports for sequential experiments. The team compared several modes: no evolution, access only to training inputs, self-learning, few-shot learning with reference answers, and workflow synthesis after three rounds of reviewer feedback. Using Claude Code with Opus 4.8, for example, baseline accuracy was 49.11%; self-learning raised it to 57.37%, while three rounds of reflection reached 62.72%. Providing example answers directly produced 70.90%. Gains from autonomous evolution varied across models and execution frameworks, showing how difficult it is to measure “model capability” separately from the harness—including memory, tools, and prompt composition.
The main technical value is that the data pipeline can rapidly generate new task sets. The authors say they expanded V1 from 120 to 240 tasks in two days, allowing evaluators to rotate questions and reduce contamination risk. Engineering teams should still not treat the benchmark as a measure of enterprise deployment success: the tasks are generated by an automated pipeline, scoring is largely rule-based, and the benchmark does not yet cover real-world organizational constraints such as permissions, ambiguous requirements, and long-term state drift. The next step is to see whether other teams can reproduce the results and whether the strategies accumulated by agents remain effective across task sets, industries, and version updates.