# Performance Drift in Agent Systems

Canonical: https://metix.ai/research/agent-performance-drift

Engineering note on why production agent systems drift across prompt, architecture, evaluation, and context layers, and how to manage that drift.
Agent drift is structural, not incidental: teams need prompt versioning, layered evaluation, golden sets, observability, context engineering, and deterministic fallbacks.
## FAQ

**What is performance drift in agent systems?**

Drift is the gradual shift of a production agent's behavior across four layers, prompt, architecture, evaluation, and context, even when the specification and the business goal stay fixed. A model upgrade alone can change tool-use ordering or break output parsing that months of prompt work had stabilized.

**Why can evaluation scores go up while user complaints also go up?**

Because the ruler itself bends. Development-time metrics often stay parked on engineering-friendly measures like output correctness, while the real success criteria in production are retention, task completion, unit cost and complaint rate. The evaluation system typically lags the business understanding by six to twelve months.

**What makes a good golden set?**

It does not have to be large: 100 to 500 cases. It must be human-annotated or human-adjudicated, frozen for at least six months, distributed close to production traffic, and kept out of any training or fine-tuning pipeline to avoid contamination. It is the absolute reference other evaluations are compared against.