Intent Drift
Intent Drift describes how agent work gradually separates from the original human objective, constraints, and accepted outcome criteria.
Intent Drift is the gradual separation between the original human objective and the direction an agent system actually follows. The words in the current prompt may still look reasonable, but the work has moved away from the intended outcome, active constraints, or strategic judgment that gave the task meaning. Intent Drift is not only a model misunderstanding. It is a lifecycle failure where the system lacks a stable way to preserve, update, and verify intent over time.
The problem it names
The problem it names is quiet misalignment during long-running work. A user may begin with a clear objective, clarify several boundaries, reject certain paths, and approve a direction. After enough sessions, summaries, tool runs, and handoffs, the agent may continue with a plausible but altered interpretation. No single step looks obviously wrong. The drift appears when the final direction no longer matches the reason the work existed.
Why existing approaches are not enough
Better prompts reduce ambiguity at the moment of instruction, but they do not guarantee that intent survives change. Longer context windows can hold more material, but they do not decide which intent is current, which decisions are historical, and which constraints still bind execution. Logs can show actions after the fact, but they do not prove that each action stayed attached to the approved objective. Intent needs to be represented as active lifecycle state.
How it relates to AI Agent Lifecycle
Within AI Agent Lifecycle, Intent Drift is one of the core failure modes the lifecycle is designed to detect and reduce. The lifecycle asks for a traceable path from intent to context, plan, confirmation, execution, evidence, review, and accepted outcome. If that path cannot be reconstructed, drift becomes invisible until a human notices the mismatch. Protocol Engineering makes the intent object explicit; Evidence Chain makes the adherence inspectable.
White paper source trace
Intent Drift is adjacent to GAIC through lifecycle continuity; R3K-0 did not assign a direct chapter/table/MRO anchor for this route.
The concept supports lifecycle governance by naming how work can separate from the original purpose before acceptance.
A plan can remain executable while no longer satisfying the intent, constraints, or acceptance state that made the work legitimate.
This source trace is author-analytical. It is not legal advice, certification, legal compliance proof, regulator approval, vendor ranking, procurement guidance, or a claim that MPLP is required.
Evidence route
The evidence route for Intent Drift runs through the origin essay and the lifecycle governance essay. The origin essay names the experience of project intent shifting across prompts, summaries, and sessions. The governance essay then shows why intent needs a lifecycle authority layer rather than a one-time instruction. MPLP, Cognitive OS, and Validation Lab each test a different part of that route: protocol, runtime state, and evidence review.