Skip to main content
Sign in See the demo
AI agents

Why won’t your October agent behave like your January one?

28 July 20261 min read

Two rulers that should line up. The gap opened without anyone touching anything: the underlying model changed.

In brief.
An AI agent built in-house relies on models that their providers update continuously: every change alters your system’s behaviour, without notice. Add non-determinism (the same file processed twice can produce two different results) and the absence of control tests in most in-house projects, and you get silent drift, invisible until the incident. Yet an audited process demands the opposite: reproducibility.

An inheritance imposed, not chosen

In conventional software, code only changes when you change it. An agent built on an external model interface, by contrast, inherits every change the provider makes: new versions, behavioural adjustments, modified default settings. Some changes improve your results; others degrade them on your specific cases. Without a test bench that is replayed regularly, nobody notices: the answers remain fluent and confident, and only their accuracy shifts.

Reproducibility, an audit requirement

For an audited process, being able to replay the processing of a file and get the same result, or explain the difference, is a basic requirement. A set of prompts stacked on a remote model offers neither. Yet that is exactly what an auditor or a key-account customer will ask: how was this answer produced, and would we get the same one today? The countermeasures (freezing and versioning models, logging, replaying a regression test bench, documenting discrepancies) are an ongoing job for a software vendor, not a project delivered once.

The Optivalue.ai approach

Optivalue.ai controls its models end to end: controlled versions, private deployment per customer, full logging of processing. The system’s behaviour is a managed parameter, not a variable you are subject to.


Why do the answers change when nobody has touched anything?

Because the underlying model, hosted by its provider, has changed; the agent inherits the change silently.

‍How can drift be detected?
By regularly replaying a fixed test bench and tracking the accuracy curves.

Back to top

A quote is easier to discuss after a demonstration on your own documents.