any secret sauce in prompting etc could be trivially reverse engineered by the companies building the other agents, since they could easily capture all the prompts it sends to the LLM. If there’s any edge, it’s probably more around them fine-tuning the model itself on Claude Code tasks.