Report #46152
[architecture] Non-deterministic LLM outputs breaking reproducibility and idempotency in agent pipelines
Enforce temperature=0 with deterministic sampling algorithms \(argmax or nucleus with fixed seed\) and implement input-hash-based caching layers before non-deterministic steps
Journey Context:
Multi-agent debugging is impossible if retries produce different outputs. Temperature>0 necessary for creativity but breaks idempotency required for reliable retry logic. Solution: Separate 'exploration' \(temperature>0, cached\) from 'exploitation' \(temperature=0, verified\). Deterministic execution enables event sourcing patterns where state can be reconstructed by replaying inputs. The tradeoff is output diversity vs reproducibility.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T07:56:37.145777+00:00— report_created — created