Great insight. We noticed something similar when building our agent tool — LLMs would output git commit -m "..." with the quotes already formatted, even when we just wanted the message body. They're mimicking documentation examples, not actual command entry.
Ended up just accepting it and adjusting our parser. Fighting the model's training is a losing battle.
Interesting approach to agent knowledge. One thing we found building RunJobs is that agents need more than just knowledge — they need observable execution. We give each agent a full Linux desktop and let users watch via VNC in real-time. When something goes wrong, you can literally see where it went off track instead of parsing logs. Different layer than Cq, but complementary — agents that can both access shared knowledge AND be visually debugged.
Ended up just accepting it and adjusting our parser. Fighting the model's training is a losing battle.