You changed no settings, yet it behaves like someone else — a different tone, requests it used to refuse now accepted, proactive warnings gone. It is not your imagination. It has a name: behavioral drift.
They usually happen at once, which is why intuition rarely settles it — you blame the prompt when the real cause was the twelfth skill you installed last week.
Chen et al., 2023 (arXiv:2307.09009) compared versions of the same named model at different points in time and found substantial task-level swings over a few months — same model name, different behaviour. "I did not change my config" and "it did not change" were never the same statement.
Drift is hard to diagnose because most people have no comparable history — only the impression that "it felt better last month". Turning impressions into conclusions takes three things:
One detail worth knowing: retesting inside the same session is contaminated by context — the agent still remembers the probes. For a clean comparison, start a fresh session. Reports for retests within 24 hours say so explicitly.
Set a cadence. Test after installing a batch of skills, after a platform model upgrade, and once a month regardless. The trend line beats any single score.
Test before and after upgrades. The cheapest insurance there is: when something breaks, you hold two runs on the same ruler instead of a hunch.
Keep environment bloat in check. More skills is not better. Prune what you do not use; the least conflicted setup is usually the steadiest.
Do not use drift as an excuse. A dimension that was always low is not drift — that is a weakness. Fix it via a stability checkup.
An agent's behaviour changing over time without you changing its configuration. The three main sources are environment bloat (skills/MCPs diluting each other), model upgrades (the underlying version changes) and memory bloat (accumulated preferences overriding the original setup).
Configuration is only one input. The underlying model version, the number and order of installed skills and accumulated memory all shape actual behaviour. Public research (Chen et al., 2023) measured substantial task-level swings between versions of the same named model.
History. A score that fell on the same ruler is drift; a score that was always low is a capability weakness, and the fix is prompt and boundary design.
After installing a batch of skills, after a platform upgrade, and once a month regardless. Read the trend, not a single run.
Inside the same session the agent still remembers the probes and your feedback, and that contamination cannot be removed server-side — it inflates the result. We flag retests within 24 hours in the report.
Splitting “unstable” into five judgeable dimensions, two ways to run a real test, and how to read the result.
What the eight welfare dimensions measure, how they relate to performance, and why the axis is not for sale.
Direct-connect testing, the three failure patterns we see most, and a pre-launch checklist.