QA Web Agent

Your agents write the code.
Who checks it?

Coding agents open pull requests faster than any team can click through them. QA Web Agent puts an autonomous agent in a real browser against a real copy of your app, posts the verdict on the pull request, and when the bug is confirmed, writes the fix and hands you the diff.

The Pull requests view: a watched PR with its verdict label, the planned suite and per-test results the agent posted to GitHub.

Most QA tools test the URL you hand them and stop at a report. This one builds the environment, proves the bug in a browser, and closes the loop with a change you can merge.

Watcha PR, an issue, a schedule
Runa real browser, a real app copy
Verdictlabel + comment on GitHub
Filerepro steps, video, screenshots
Fixagent change, self-validated
Graduatea reviewed PR on your repo

It uses the product like a person.

A test is a paragraph of plain language. The agent finds its own way past sign-in and setup, retries once before it ever says fail, and leaves screenshots, video and a transcript behind for whoever asks "are you sure?"

Regression tests

It answers where reviewers already are.

One verdict label and one comment on the pull request, updated on every push against the exact commit. Issues get reproduced, retested and marked verified. Failed runs become filed bugs with steps and a recording.

Pull requests

It ships the fix, not the ticket.

Assign a confirmed bug to the code agent. It edits, boots your app from its branch, runs the same test that proved the bug, and a reviewer graduates the change into a real pull request. Nothing merges on its own.

Agent code changes

Built for the era of malleable software.

Users and agents now reshape software at the point of use. Feedback no longer ends at "we passed it to the team." A vendor plugs its product into a project of its own, with its own repository, environment, data and model. A customer's report becomes a reproduced bug, a validated change, and a pull request on the vendor's repository, watched until it merges.

This is not a slide. It has run end to end on Jade, an event-housing product with seventy agent tests, four MCP servers of its own, and a feedback wizard whose issues are handed straight to the code agent.

How vendors plug in

A code change under review: branch, commit, validated status, the brief, and the Graduate to PR action.

Sixteen features. Five things they do.

Test

Plain-language tests, exploration, and the environments they run in.

Verify

Verdicts where the work happens: pull requests, issues, evidence, trends.

Fix

From a confirmed finding to a reviewed change.

Document

Requirements and release notes that follow the code.

Run

Schedules, webhooks, models, integrations, many products on one platform.

The Model view: OpenAI, Anthropic, Hermes or the Claude agent per project, with a fallback harness.

An open harness, not a black box.

Choose the model and the harness per project: OpenAI or Anthropic through the built-in loop, the Hermes repo-aware harness, or Claude Code on the Claude subscription you already pay for. Hand the agent your app's own MCP tools. Drive the whole platform as an MCP server. Every job records what it cost.

Models, MCP and API

We test it with itself.

Every pull request to this platform is watched by this platform: the agents clone the PR, boot a throwaway copy of the whole system, run the dashboard suite, sweep the screens for accessibility and layout, regenerate the requirements docs and draft the changelog. The site you are reading was tested that way before it shipped, and the findings were fixed before it merged.

How it works, end to end

The dashboard's Home canvas mapping the QA lifecycle to what the platform automates today.

Questions people ask

Does the agent write to our repository?

Not on its own. Agent-made changes live on a git host the platform owns until a reviewer graduates one; only then does the platform open a real branch and pull request on your repository, and it never merges.

Which AI models and harnesses can drive the agent?

OpenAI and Anthropic models through the built-in loop, the Hermes repo-aware harness, or Claude Code signed in on your own Claude subscription. The choice is per project and needs no redeploy.

Where do results show up?

On the pull request as a verdict label and one updating comment, on the issue as a retest comment, in Slack or any webhook, over a live event stream, and in the dashboard with screenshots, video and the full agent transcript.

Is this a replacement for a QA team?

It is an accelerator. The agent takes the repetitive verification, reproduction and regression work; people keep the judgement calls: what to watch, which findings matter, and whether a proposed change ships.

What does a run cost?

Every job records its model, tokens and dollar cost, so spend is visible per test, per suite and per project rather than hidden in a seat price.