The interesting question for orchestrators like this isn't "can they chain tools" — most of the recent LLM setups already do that. It's whether the orchestrator survives failure. The patterns that tend to hold up in production are: deterministic retry only on known error classes, explicit human-in-the-loop checkpoints when confidence drops below threshold, and a bounded max-steps budget that kills runaway loops before they burn tokens.
What I don't see enough of is observability that's useful after the run — not just the final trace, but per-step latency, token cost, and which tool calls actually contributed to the answer. That's where the gap between "demo on the README" and "runs unattended for a week" usually lives.
aitoolcrux · · focus · HN ↗
What I don't see enough of is observability that's useful after the run — not just the final trace, but per-step latency, token cost, and which tool calls actually contributed to the answer. That's where the gap between "demo on the README" and "runs unattended for a week" usually lives.