Card 44: Agentic MCP → CI

The portable agent skill for this pattern is playwright-agentic-testing.

What This Pattern Solves

A browser agent can pursue a goal through Playwright MCP. That is useful for exploration, flake debugging, and reproducing bugs you cannot fully script yet.

It is a bad fit for high-frequency CI. The durable handoff is: explore with MCP, review the path, commit a deterministic Playwright test, run that in CI. Leave the live agent out of the pipeline.

This card is the committed half of that loop. Card 42 (WebMCP) is different: the page exposes tools to agents. Here the agent drives the browser, then you keep a normal test.

Goal an agent might pursue

Open /login. Sign in as testuser / password. Land on the dashboard and confirm
the greeting mentions testuser. Prefer roles and labels over CSS selectors.

How It Works

1. MCP explores; CI executes. The agent session discovers the path. The committed file is the regression gate.
2. Review before you keep it. Rewrite brittle drafts with roles, labels, and web-first assertions (playwright-architecture, playwright-locators, playwright-assertions).
3. Do not put the agent loop in CI. Known paths belong in Playwright tests.

When To Use

• You used Playwright MCP to find or debug a UI path and want a lasting check.
• You generate a first draft with an agent and need a pattern for what “done” looks like in the suite.
• Not when you are testing tools the page registers for agents — that is Card 42.
• Not when a cheaper tier already owns the claim.