Behavioral drift is the term From Determinism to Delegation: AI-Native Software Engineering and the Evolution of the Agentic Engineer uses for the way agentic systems degrade, which it sets against deterministic software as the thing that does not. The paper separates two distributional shifts. Data drift is a shift in the marginal input distribution: the inputs the system meets in production stop resembling those it was trained on. Concept drift is a shift in the conditional: the same input should now produce a different correct output. The paper states the two separately without commenting on how they relate.
The paper names a further cause alongside those two. Because the underlying model may also be updated by its provider, an agent can change behaviour without anything in its own deployment changing. The consequence the paper draws is that an agent has no stable terminal state and requires permanent stewardship together with automated statistical monitoring under testing, evaluation, verification and validation regimes.
Usage
The concept is used to argue that “done” is not a meaningful state for an agentic system. In that paper’s fourteen-dimension comparison it is what the lifecycle row turns on: classical practice runs spec, design, implement, test and deploy toward a clear “done”, while agentic practice is experiment-heavy and trace-driven with no terminal “done”, because behaviour drifts.
It is worth reading beside Compositional Reliability, which the same paper treats in the adjacent section: that one concerns how far an agent can be trusted within a single long run, this one how long any measurement of it stays valid. The paper discusses both among the mechanisms underlying autonomous software agents; the pairing here is this wiki’s, not a grouping the paper makes.
When It Applies
The paper states the contrast as agentic systems degrading through drift where deterministic software does not. The provider-update case widens that further: it applies even where the deploying organisation changes nothing at all, which is what makes it a governance concern rather than only an engineering one.
What it demands, in the paper’s terms, is permanent stewardship and automated statistical monitoring rather than a test suite that passes or fails, because the thing being watched is a distribution rather than a value. The paper connects this to its treatment of evaluation as the central artifact of agentic practice, and specifically to robustness being quantified through frameworks that exceed static accuracy and evaluate reaction to data, model and intent drift.
Its consequences reach the paper’s open problems from two directions. Under sustained stewardship, it states that the cost model of permanent monitoring, re-evaluation and re-alignment is not yet well understood at the portfolio level. Under accountability attribution, it names vendor-supplied model updates that silently alter behaviour as part of what makes responsibility allocation legally and ethically unsettled — the drift is not merely a reliability problem but a question of who is answerable when a system changes underneath its operator.
This account rests on one position paper, which states the two distributional shifts and the provider-update cause compactly and draws its conclusion from them jointly. It reports no measurement of drift of its own.
Related Terms
- From Determinism to Delegation: AI-Native Software Engineering and the Evolution of the Agentic Engineer — the paper this account is drawn from
- Compositional Reliability — the adjacent mechanism the same paper treats, on a different time axis
- Context rot — a distinct within-session degradation, not this cross-time one
- AI-Native Software Engineering — the paradigm in which permanent stewardship replaces a terminal “done”