When a test fails, Currents now tells you why. Open the test and the Insights panel has the likely cause of the failure and suggested fixes, written by AI. When the test has a Playwright trace, the AI reads it: it goes through the actions, opens the network requests around the failing step, and looks at the page when the test failed.

AI analysis is available to every organization, on every plan. An organization admin turns it on once in the organization settings, under AI and data.

A Playwright error says what failed but not why. The AI reads the Playwright trace, step by step, and writes the likely cause with a suggested fix. On 156 real CI failures it found the right cause 81% of the time, against 6% from the Playwright output alone.
The likely cause of a failed test

The likely cause of the failure

Expand Likely cause of the failure on a failed test. While the AI works, the card lists what it checks: the actions it reads, the requests it opens, the screenshots it looks at. In our tests it took about 20 seconds. When it finishes, the card shows:

  • The category of the failure: a network error, an element the test couldn't find or click, an assertion, a timeout, a config or a code error.
  • The root cause, in a few sentences.
  • Suggested fixes you can try.

The result stays with the test, so the next person who opens it sees it right away. Copy it into the pull request, the issue, or a prompt for your coding agent.

How we tested it

A wrong explanation is worse than none: it sends you after the wrong cause. So before turning this on for everyone, we measured it.

We took 156 real CI failures from 53 runs of a production Playwright suite, with the customer's permission. Independent investigator agents, using a different AI vendor, worked out the real cause of each one from the full trace. Then we ran each version of the analysis on the same failures and graded every answer against those causes.

The Playwright error output alone rarely explains a failure
  • The first version read only the error, the steps and the console output. It found the right cause 6% of the time. Most answers hedged between two guesses.
  • Adding the evidence around the failing step (network requests, the page snapshot, timing) helped most when the cause was a backend error.
  • Letting the model investigate the trace itself, with tools to list actions, read requests and response bodies, and look at screenshots, helped the most.
  • We compared five models from two vendors. The cheapest model got to 59%. The one we use, gpt-6.1-terra, at medium effort, got to 76%.
  • Making the trace tools easier for the model to read (response shapes, request timing, flagging a stale page snapshot) took it to 79–81%, depending on the grader. On the failures where the cause was certain, it was right 88% of the time.
Each step, measured on the same failures

The Playwright-output-only answers came from the smaller model. In an earlier grading round we gave gpt-6.1-terra the same Playwright output, and it was right 5% of the time. A stronger model doesn't help without the trace.

What didn't help: the test's history, its source code, longer instructions, and more reasoning effort. High effort was no better than medium, and twice as slow.

The limits, so you can judge the numbers. All 156 failures come from one project, so we didn't tune the prompt for it. Two independent investigations agreed on the cause 77% of the time overall, and every time when both were certain, so numbers above that are hard to measure. There was no fine-tuning: no model was trained, on your data or anyone else's. The work was choosing what the model sees, which tools it gets, and which model to use.

When a test has no trace, or reading it fails, the analysis uses the error, steps and console output alone. Expect less from those answers.

What does this test check?

Explain this test works on any test, passed or failed. It describes in plain words what the test does and what it checks. Useful when you review a failure in a test you didn't write, or the test name doesn't say much.

Turn it on

AI analysis sends test data to an AI provider, so it needs your organization's consent:

  1. Open the organization settings and select AI and data.
  2. Turn on AI assistance.

Only admins can change it. Members who open a test before then see who to ask.

When you want an agent to fix the failure in your code, connect it to Currents with the MCP server. It reads the same failure data and has your repository too.