Review the evidence, not the diff.
Agents open more pull requests than your team can read. Currents links the evidence from every Playwright run to each one. Your agents fix what they broke. You decide what merges.
feat(checkout): one-tap pay with a saved card #482
agent wants to merge 8 commits intomainopened 02:14
41 files · 1,596 lines · about 4 hours to read
41 files changed, with 1,284 lines added and 312 removed.
Viewed 2 of 41 files
About 4 hours to go at 400 lines an hour
Currents Tests1,204 passed · 0 flaky · 6m 12s
RequiredPassedcheckout › pays with a saved card
Last 30 days
29 of the last 30 runs passed. The one failure was Run #1 at 02:15, before the fix.
agent commented at 02:31
The saved card was charged before its token loaded. Fixed in pay.ts.
The fix is three lines in src/checkout/pay.ts: two added, one removed.
Touches payments, so it waits for a human read.
Review required · 11 files, about 45 minutes
Code got cheap. Knowing it works didn't.
Stripe merges over 1,300 pull requests a week with no human-written code. A year after companies adopt coding agents, pull requests per engineer are up 36%, and review takes 56% longer.
The work of checking code didn't go away. It moved.
An agent checking its own work isn't verification
The checking can't move to the agent that wrote the code. An agent can make a red test pass without fixing anything. To tell a real failure from a flaky one, it needs a record it didn't write.
Agent alone
agent~/shop
$gh run view --log-failed
└pays with a saved card
Probably a slow page. Adding a wait.
tests/checkout/payment.spec.ts+1
+await page.waitForTimeout(5000);
$npx playwright test payment.spec.ts
└1 passed
- It sees one local run and a CI log, so it guesses.
- It adds a wait or loosens an assertion until the test passes.
- The check goes green. The bug stays.
Agent + CI evidence
agent~/shop
currents-get-test-results
└29 of last 30 runs passed, 0 flaky
currents-get-context
└error and trace from Run #1 of #482
It has never flaked. The failure is real.
Fixing src/checkout/pay.ts, not the test.
- It reads the test's history across runs: 29 of the last 30 passed, none flaky.
- It reads the error and the trace that CI recorded.
- That's evidence the agent didn't write, so it fixes the code and leaves the test alone.
From reading code to reading evidence
Verification engineering is the practice of proving that a change works with evidence from tests that run in CI, instead of a person reading every line. Before agents, that reading was the bottleneck. With agents, the bottleneck is CI running the tests and someone reading the evidence. People keep both ends: deciding what to build and deciding what merges.
Sumit Saurabh explains why the work moved and how Currents handles it.
Before agents
- Decide what to build
- A person writes the code
- The old bottleneckA person reads every line
- Merge
With agents
- You decide what to build
- An agent writes the code
- The new bottleneckCI runs the tests, someone reads the evidence
- You decide what merges
Fails? The agent fixes it and pushes again.
From red build to proven fix
The agent's pull request #482 fails CI. It checks whether the failure is real, fixes it and puts the proof on the pull request.
CI fails on the agent's pull request
feat(checkout): one-tap pay with a saved card#482
Currents TestsRequired
Run #1 failed · Merging is blocked
One test fails on Run #1: checkout › pays with a saved card. The Currents check is required, so merging is blocked.
The agent checks history, then reads the trace
checkout › pays with a saved card
currents-get-test-results
29 of last 30 runs passed, 0 flaky
The test hasn't flaked in 30 days, so the failure is real. The agent reads the error and the trace over MCP and the Trace API: about 1,600 tokens for that one trace instead of about 20,000.
It fixes three lines and re-runs
src/checkout/pay.ts+2−1
−const token = tokenForSavedCard(card);
+const token = await tokenForSavedCard(card);
+if (!token) throw new PaymentError("card_token_missing");
re-run · 1 passed (6.2s)
The saved card was charged before its token loaded. The agent adds the missing await in
src/checkout/pay.tsand re-runs the test: 1 passed.It posts the before and after on the pull request
Run #1's declined card sits beside Run #2's confirmed order. Reviewers open both without a Currents login, until the link expires.
feat(checkout): one-tap pay with a saved card#482Openagent wants to merge 8 commits into mainCurrents Tests · 1,204 passed · Run #2Requiredagentcommented nowThe saved card was charged before its token loaded. Fixed in pay.ts.Before and after, from Run #1 and Run #2:✕ Before · Run #100:06 / 00:061×✓ After · Run #2ScreenshotTrace:Run #1· opens without a loginPut evidence on every agent PR14-day free trial. No credit card required.
You decide what merges
11of #482's 41 files are yours to read
Payments, the migration and the tests. About 45 minutes instead of 4 hours.
The other 30 merge on 1,204 passing tests.
You set which files people read and what the rest must pass. Each rule lives in one place.
People read what's risky
/src/checkout/pay.ts@acme/payments
/src/api/payments/@acme/payments
/src/db/@acme/platform
/tests/@acme/qa
# everything else merges on evidence
The paths you name wait for a review from their owners.
Set in CODEOWNERS and branch protection · GitHub
Red can't merge
Currents TestsRequiredRunning 1,204 testsMerging is blockedMerge pull requestA failing Currents check blocks the merge button, for agents and people alike.
Set in Branch protection · GitHub
Flaky tests don't decide
search › autocompleteQuarantinedKnown flake · expires in 14 daysA quarantined test still runs and its failure doesn't fail the run. Give a rule an expiry date, and Currents alerts Slack or email before it lapses.
Set in Currents Actions
Agents get what you approve
Claude CodeAsks to read resultsCan read resultsApproveRevokeAn agent's access is capped by the role of the person who connects it, and you can revoke it.
Set in Connected Applications · Currents
For your agents
The Playwright Best Practices skill teaches agents locators, web-first assertions and debugging flaky tests. Installs from GitHub, no Currents account needed.
Playwright Best Practices skillFor your security review
- SOC 2 Type II
- SAML SSO and SCIM · Business
- Data redaction · Enterprise
- Your own storage or on-premises · Enterprise
Connected to every step, from push to merge
Each run's evidence lands where the next person or agent already works.
Runs in the CI you have: GitHub Actions, GitLab, CircleCI, Jenkins and more. CI setup docs (opens in a new tab)
People
See what happened
- GitHub
- GitLab
- Bitbucket
- Slack
- Microsoft Teams
Ship more pull requests. Read fewer diffs.
FundGuard
Read the FundGuard storyTime to a verdict
From about 80 min to 40min
FundGuard's pipeline, halved by Currents Orchestration. Its CI bill fell 70%.
n8n
Read the n8n storyTrust in tests
From about 2 to 3% to 0.3%
n8n's flaky rate, before and after its move to Playwright with Currents.
Trace API
See the launch exampleEvidence an agent can use
From about 20,000 tokens to 1,600tokens
One failing trace, pasted whole versus read through the Trace API.
“It’s easy to go in, see what’s trending as flaky, line it up with traces, and understand when and why it was introduced.”
Declan Carroll / Senior Quality Engineer at n8nThree steps to merging on evidence
Minutes from signup to your first recorded run.
Record every run
Start a trial, then add the reporter and your project's keys to CI.
Start a 14-day free trialRead the setup guidePrompt for your agentSet up this repository to report test results to Currents. Read https://docs.currents.dev/getting-started/setup-with-an-agent.md and follow it.Gate the merge
Install the GitHub App and make the Currents check required. GitLab and Bitbucket too.
GitHub App docsPull requestSome checks haven’t completed yetAll checks have passed1 queued1 successful checkCurrents TestsRequiredWaiting to run1,204 passed · 0 flaky · 6m 12sConnect your agent
Add the Currents MCP server and sign in. Your agent reads failures, history and traces.
Connect Cursor, Codex or ChatGPTTerminalclaude mcp add --transport http --scope user currents https://mcp.currents.dev/mcp
