I’ve noticed it’s the performance that suffers when agents are paired with more than a single CLI and non-prescriptive skills. Since it seems to be out of its t

I’ve noticed it’s the performance that suffers when agents are paired with more than a single CLI and non-prescriptive skills. Since it seems to be out of its training data, anything non-trivial and the model just tries to brute force its way to a solution. Maybe it’s also about building them as self-improving, though I’ve been doing it manually for a CLI we don’t own. It seems to be art at the moment.