There are a few solid approaches, layered by effort: ## 1. Eval suite against a golden dataset Capture a set of representative inputs and their expected outputs (or judged-good outputs) from when the agent was behaving correctly. Run these on a schedule or on every model/prompt change. - **Exact match** works for structured outputs (JSON, tool calls, classifications) - **LLM-as-judge** works for prose — prompt a model to score the new output against the reference - Fail the build / alert if scores drop below a threshold ## 2. Behavioral fingerprinting Track observable statistics over time, not just correctness: - Tool call frequency and order - Response length distribution - Refusal rate - Latency / token usage Sudden shifts in any of these are a canary that something changed — even if your evals are too coarse to catch it. ## 3. Shadow / replay testing Log all real production inputs. Periodically replay a sample through the current agent and compare outputs to what was logged at ship time. Catches model provider drift (e.g. OpenAI quietly updating a model) without waiting for user complaints. ## 4. Trace every run in production Use a tracing layer (LangSmith, Braintrust, Arize, Honeycomb) so you can diff a run from last Tuesday against one from today with the same input. This is your audit trail when someone says "it used to do X." ## 5. Pin what you can - Pin model versions (use `gpt-4o-2024-08-06` not `gpt-4o`) - Pin tool schemas and system prompt versions in source control - Track which version of each was live at any given time --- The practical minimum: **a golden-dataset eval that runs on every deploy + model version pinning**. The rest scales from there as the agent grows in criticality.