How to Test and Validate Agent Run Conditions
Last updated: September 23, 2026
Overview
A run condition can read perfectly and still be wrong, because it is evaluated against matters rather than against the sentence. Testing runs the condition on real matters and reports whether the agent would have fired on each one, before it is live across the firm's book.
Three problems show up this way and almost never show up in a careful reread: a condition broad enough to fire on nearly everything, a condition so specific that nothing qualifies, and folder or document names that vary from matter to matter.
For writing the condition in the first place, see 📄 How to Configure Trigger Conditions for Auto Agents. This page picks up once a condition is in the Run condition field.
Before You Start
Firm admin access.
An agent open in the builder with a run condition entered.
An idea of which matters should qualify and which should not. A test only means something against an expectation.
Testing the Condition
Click Test this condition on matters below the Run condition field to open the testing panel.
Choose the matters to test against. Select matters picks up to five specific matters from the case list, and Random pick takes a sample. Three is the working minimum.
Click Test condition.
Read the result for each matter. Condition met means the agent would have run. Condition not met means it would not have. Each result comes with an explanation of what Eve found or could not determine.
Choosing Test Matters
Which matters get picked decides how much the test is worth. Three clean files that all pass confirm nothing except that the condition works on clean files.
A useful sample of three to five:
One matter that should clearly qualify. The case the agent was built for.
One that should clearly not. A matter in the right practice area with none of the triggering activity. This is the check on breadth, and it is the one most people skip.
One messy file. Inconsistent folder names, documents uploaded in a batch, records that arrived out of order. This is where naming problems surface.
One edge case, where the answer is genuinely arguable. How Eve explains that one says more about the condition than the two obvious results do.
Random pick is a good way to sanity-check breadth across the book once the deliberate sample passes, since it will find matters nobody thought about.
Reading the Explanations
Reading the explanation can help you better understand why this item received a particular result. A result only says whether the condition fired; the explanation says why, and it usually names the fix.
A condition of "new medical records are uploaded" can come back not met because Eve could not establish what counts as new on that matter, with no prior evaluation to measure against, or because it could not tell which files arrived recently and which were already there. Those are two different problems with two different fixes, and only the explanation distinguishes them.
What the Results Usually Mean
What the test shows | Usual cause | Fix |
|---|---|---|
Condition met on every matter, including ones that should not qualify | The condition is too broad | Add a folder, a document type, or a time bound |
Condition not met anywhere, including the obvious case | A name in the condition does not match what is in the matters, or "new" is undefined | Check the exact folder and document names; state explicitly that the documents must be newly uploaded |
Met on some matters and not on others that look the same | Naming varies from matter to matter | Build the variations into the condition with AND/OR logic |
Met, but the explanation cites a document you did not intend | The condition is matching on something incidental | Narrow the scope, or exclude the document type explicitly |
Naming variation is the most common of the four. Discovery files turn up as ROGS, Interogs, and Interrogatories across three matters at the same firm, and a condition naming only one of them fires on a third of the cases it should. Write the variations in: "files named ROGS, Interogs, or Interrogatories."
Refine and Re-Test
Change one thing at a time, then run the test again. Adjusting the folder name and the time bound and the document type together makes it impossible to tell which change fixed the result, or which one broke something else.
The condition is ready when the same sample returns met on the matters that should qualify and not met on the ones that should not, twice in a row. Then save the agent and activate the trigger.