Code pattern drift detection is the use of a coding agent to compare a codebase with a reference application — a compilable, runnable project that embodies an organisation’s intended coding patterns, such as a service template — and to identify where the codebase has drifted away from those patterns, so that the gaps can then be closed. The term is used in Anchoring AI to a reference application, which describes it as a second step after using the reference application as the source of code samples a coding assistant works from.
Usage
In the setup that article describes, a Model Context Protocol server gives the agent access both to the reference application’s code samples and to its git commits. The prompt asks the agent first to find the latest changes in the reference; the agent fetches the latest commit through the MCP server, uses its diff to analyse the target application, and writes a drift report. A person reviews and edits that report — removing irrelevant findings, for instance — and then, in a second step, asks the agent to write code that closes the gaps it identifies. Scoping the comparison by the reference’s commits is how the user tells the agent what kind of drift to look for; the article reports that asking the agent simply to compare the reference and target code instead produced many irrelevant comparisons.
When It Applies
The problem it addresses is that service templates, a typical building block of the “golden paths” organisations build for their engineering teams, are meant to carry the current patterns and standards, but once a service has been created from one, feeding later template updates back into it is tedious, and generated code that is then extended and maintained often drifts from the reference. The approach assumes such a reference application exists and is maintained, and that its changes are recorded as commits an agent can read.
Where the drift is simple enough — adding a logger, changing a logging framework — deterministic codemod tools such as OpenRewrite can close it, and the article advises considering them before reaching for AI. It locates the agent’s contribution in drift whose fix needs coding more dynamic than regular-expression-based codemod recipes allow, such as turning varied, non-standardised log messages into a structured format. How well-established the technique is: it is one practitioner’s experiment, tested on a deliberately simple example, with the author reporting that commit scoping had a good impact rather than presenting any measurement.
Related Terms
Model Context Protocol, Anchoring AI to a reference application