Give your coding agents a QA loop, not just a to-do list

Coding agents ship code faster than humans can verify it. agent-qa closes the loop: through MCP tools and packaged Skills, agents author end-to-end tests from product context, run them on web and mobile, read the evidence, and fix what failed — before a human reviews the PR.

The verification gap in agentic development

Every team adopting coding agents hits the same wall: the agent says it's done, the diff looks plausible, and nobody actually exercised the feature. Unit tests catch logic errors; they don't catch a broken checkout flow. The missing piece is end-to-end verification the agent itself can run.

agent-qa was designed agent-first. Its MCP server exposes the full QA surface — project context, test authoring, validation, execution, run artifacts, and failure classification — as tools any MCP-capable agent can call. Verification happens in the same loop as the code change, not in a nightly job three merges later.

Skills that teach agents to test well

Beyond raw tools, agent-qa ships packaged Skills — reusable instructions for authoring tests, debugging failures, and triaging results — installable in one command. Your agents don't just get API access to a test runner; they get the working discipline of a QA engineer.

Agent integration
# expose agent-qa to coding agents over MCP
agent-qa mcp

# install packaged Skills for Claude Code and peers
npx skills add vostride/agent-qa --skill '*'

Shared memory between agents and humans

Because agent-qa's memory, cache, and artifacts are repo files, everything an agent learns while testing is visible to the team — and everything the team knows is available to the next agent run. Failures come back classified and evidence-backed, so an agent can distinguish 'my code broke the flow' from 'the environment flaked' and act accordingly.

tests/checkout-smoke.yaml
test-id: t_checkout-smoke
name: Guest checkout completes
target: storefront-web
context: |
  The storefront runs at the URL configured by the storefront-web target.
  Product data is seeded by the workspace setup hook.
steps:
  - Open the home page and search for "espresso grinder".
  - Open the first search result.
  - Add the product to the cart.
  - Start checkout as a guest.
  - Fill the shipping form with the workspace test address.
  - Place the order.
  - Verify the confirmation page shows an order number.

Frequently asked questions

How do coding agents run E2E tests with agent-qa?

Start the MCP server with agent-qa mcp and register it with your agent (Claude Code, Cursor, or any MCP-capable client). The agent then calls agent-qa's tools to author tests, validate them, execute runs, fetch artifacts, and classify failures — all within its normal working loop.

What are agent-qa Skills?

Packaged, versioned instructions that teach agents specific QA workflows — authoring tests from product context, debugging and fixing failures, and triaging results. Install them with npx skills add vostride/agent-qa.

Can agents and humans share the same test suite?

Yes — that's the point. Tests are plain-English YAML in the repo, so a test written by an agent is reviewed by humans in a PR, and vice versa. Memory and artifacts accumulate in the same repo-visible locations regardless of who triggered the run.

Why not just let the agent drive a browser directly?

Raw browser control (for example via a browser MCP) gives an agent hands, not judgment. agent-qa adds the harness: durable test contracts, deterministic environment handling, hooks, memory, caching, and classified evidence — the difference between an agent that clicked around and an agent that verified.

Keep reading