Skip to main content
Momentic uses a variety of agents, each specialized for a specific task within end-to-end testing. Currently, the following agents can be configured by users:
NameDescription
Locator (locator)Locates elements on the page based on a natural language description. This agent is used for all interactive commands that target elements, such as Click and Type steps. This agent is also used to resolve elements in Element check steps.
Assertion (assertion)Evaluates the truthiness of natural language statements based on an instantaneous snapshot of the page. This agent is used to power AI check steps.
Visual Assertion (visual-assertion)Evaluates the truthiness of natural language statements purely based on a screenshot of the current viewport.
Text Extraction (text-extraction)Extracts text from the page based on a natural language description. This agent powers AI extract steps.
Momentic exposes the ability to choose the versions of these agents so that users can upgrade at their own cadence and test upgrades before applying them.

​
Locator

​
v1

  • Uses older 2024 models.
  • Interprets the user’s criteria more leniently and will attempt to find a candidate element that fulfills the same function as what the user describes.

​
v2

  • Uses early 2025 models.
  • Places a strong emphasis on respecting relative positioning criteria (e.g. “the link in the header”, “the X icon in the modal”), and will fail if it cannot find the reference element.
  • Strongly respects single-quoted text (e.g. “the ‘Submit’ button”) and will reject candidate elements that do not contain that exact text.
  • Recommended configuration for new organizations.
  • Uses the latest 2025 models.
  • Generates smarter caches:
    • Element attributes: if a prompt specifically refers to certain attributes such as text content or position, those will be stored and required to remain the same across runs.
    • Relative elements: if a prompt refers to other elements on the page, those will also be included in the cache to ensure that any meaningful changes result in the cache busting.
  • Writes better explanations of its reasoning for choosing elements.

​
Assertion

​
v1

  • Uses older 2024 models.
  • Interprets assertion criteria more leniently and will attempt to evaluate whether or not an assertion is “overall true” or “overall false”.

​
v2

  • Uses early 2025 models.
  • Displays superior understanding of sorted order, visual positioning, and relative positioning criteria (e.g. “The button in the modal is enabled”).
  • Strongly respects single-quoted text (e.g. “element with title=‘hello world’”) and will fail the assertion if there is no element that contains that provided text.
  • Recommended configuration for new organizations.
  • Uses the latest 2025 models.

​
Visual Assertion

​
v1

  • Uses older 2024 models.
  • Faster, returning results in an average of 3 seconds.
  • More lenient when evaluating user criteria against the page state, specifically for assertions about color.

​
v2

  • Uses early 2025 models.
  • Slower than V1 but perceives small visual details more accurately.
  • Far stronger on logical reasoning, negated assertions, and assertions about elements in containers.
  • Strongly respects the term “exactly” and single quotes around text terms. Will fail the assertion if it not unambiguously true when these modifiers are provided.
  • Recommended configuration for new organizations.
  • Uses the latest 2025 models.

​
Text Extraction

​
v1

  • Uses older 2024 models.
  • Recommended configuration for new organizations.
  • Uses the latest 2025 models.
  • Adheres to the provided JSON schema more consistently and will throw an error if the extracted data does not conform to the schema.
  • Tuned to understand the pattern JSON schema option specifically.
⌘I