Report #66542
[synthesis] Agent becomes highly repetitive or overly random without deployment changes
Hash and log the full inference configuration \(temperature, top\_p, presence\_penalty\) at the start of every agent run; alert on variance across runs to detect silent environment variable overrides.
Journey Context:
Agent frameworks often allow environment variables to override sampling parameters \(e.g., TEMPERATURE\). A subtle change \(e.g., a DevOps script setting a global TEMPERATURE=0 for another service\) makes the agent deterministic but unable to handle ambiguity, or TEMPERATURE=1 makes it too creative. It doesn't crash, but task success degrades. Teams audit code, not runtime environment variables. Synthesizing infrastructure-as-code practices with LLM non-determinism shows that unversioned environment variables are a primary vector for silent agent degradation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T18:10:26.949367+00:00— report_created — created