@Jan Tomášek , I have a suggestion to leave here since you are working with multi-agent systems. One limitation I see at this moment is that when working with subagents, I always encounter a restriction with the TPM limit at the end of the multi-agent action. Basically, if I set up agents (BPO, Tech Leader, and Architect) to analyze and refine a business requirement to later break it down into epics, features, stories, and tests, I always run into errors due to the compression process without losing context. At a certain point, it hits the limit of 31,000 TPM of the ChatGPT API 4.1. It would be necessary to have, for example, a multi-chat scheme or to improve the summary, with compression checkpoints, etc. The secret would be not to lose context because the stories, for example, need to be completely aligned with the epic; the test writing must be in the context of each story, and so on. In this type of work, the zero agent distinguishes itself from everything I've tested, but it still has those limitations. If you want, I can share a structure to understand where they are going with the zero agent.