Searches for “playwright” jumped 450% this week among people also searching for n8n — the second-fastest rising related query for n8n in the US over the past seven days (Google Trends Explore, US, Nov 4–11, 2026; the rising lists shifted between page loads that morning, so treat the ranking as a signal, not a stable leaderboard). The trigger is a single release: Playwright 1.65.0, published November 5, 2026.
If you run browser automation anywhere near n8n — scraping, form filling, testing the app your workflow drives — this release matters twice. First, it ships eight breaking changes that can silently break existing suites. Second, it lands on top of two earlier releases that quietly turned Playwright from a test framework into the browser your AI agent can drive. This article separates what 1.65 actually ships from what other coverage wrongly credits it with, lists the eight breaks with fixes, and shows the n8n wiring pattern for the agent-era browser stack.
What Playwright 1.65 actually ships
The headline feature is Bun support. The test runner now runs under Bun — bunx --bun playwright test — with Bun transpiling your TypeScript test files itself instead of Playwright using Babel and its compilation cache. Playwright’s own CI runs the test runner under Bun 1.4.3 on Linux. The limits are real and documented: no watch mode, --only-changed does not follow imports, custom ESM loaders are not supported, stack traces lose async frames, and console output does not reach the trace or UI mode. If your suite leans on any of those, Bun mode is a speed experiment, not a migration.
The other additions are smaller but useful for anyone running agents against a suite:
- Reports with only failures. A new
failuresreporter prints failures and successful retries as they occur, followed by a summary, with no progress output. TheonlyFailuresoption does the same filtering for the HTML, JSON, and JUnit reporters — less noise for an agent triaging a red build. - Exclusive test locks. A built-in
-exclusive-lock runs a test only when no other test is running, for tests that touch a shared resource like a global account setting. Everything else keeps running in parallel. test.skip(),test.fixme(),test.fail(), andtest.slow()now accept only a description, recorded as a test annotation.- New APIs:
timeoutandsignaloptions incdpSession.send(), andtestRun.mode(), which returns the execution mode of a suite duringreporter.preprocess(). - New environment variable
PLAYWRIGHT_TRANSFORM_CACHE_DIRfor the transpiled test files, useful on shared Linux machines.
Browser versions in 1.65: Chromium 158.0.8127.2, Firefox 159.0, WebKit 27.2, tested against Chrome 157 and Edge 156. Node.js 20 is no longer tested — use Node.js 22, 24, or 26 — and the Docker images now ship Node.js 26 instead of 24. If your n8n instance runs Playwright in a container or through an Execute Command node, that Node bump is the first thing to check.
A correction before the checklist: the agent features are not 1.65’s
Several write-ups of this release — including a breaking-changes guide published November 8 — describe browser.bind(), page.screencast, and the CLI debugger for agents as 1.65 headline features. They are not. Playwright’s own release notes place them in version 1.59:
page.screencast— video recording, real-time frame streaming to a vision model, overlays, and “agentic video receipts”: an agent records a walkthrough video with chapters and action annotations as evidence of its work.browser.bind()— binds a launched browser to a named session so other clients can connect:playwright-cli attach my-session,@playwright/mcp --endpoint=my-session, orchromium.connect(endpoint)from another Playwright client. Multiple clients at a time are supported;browser.unbind()stops accepting new connections.playwright-cli show— a dashboard listing bound browsers, their statuses, and live sessions, with a button to open DevTools on a background browser. Playwright CLI binds all of its browsers automatically.- CLI debugger for agents —
npx playwright test --debug=clilets a coding agent attach and debug tests overplaywright-cli, stepping through a failing test instead of re-running it blind. - CLI trace analysis —
npx playwright traceexplores a trace from the command line, so an agent can read the failing action and its snapshot without opening the Trace Viewer.
Version 1.64 added the next layer: WebMCP, via page.webmcp and frame.webmcp, which exposes the tools a page registers through the experimental WebMCP browser API — testable like any other part of your app, and offered to agents as webmcp_<tool> through Playwright MCP.
Why does the version attribution matter? Because if you upgrade to 1.65 expecting a new agent feature and your suite breaks, you will blame the wrong thing. The agent browser already exists — it shipped in 1.59 and 1.64. What 1.65 does is change the ground under it. That is the actual story, and it is the reason n8n builders are searching: Playwright is no longer just the tool that tests the app; it is the browser the agent drives, and the upgrade that touches it has eight sharp edges.
The eight breaking changes, with fixes
All eight come from Playwright’s official 1.65 release notes. Three of them will bite browser-automation suites first.
1. Page-object fields that read this.page now throw. Playwright now compiles class fields the way tsc does with useDefineForClassFields, which is on for ES2022 and later targets — and when tsconfig.json sets no target at all. Field initializers run before the constructor assigns this.page, so readonly submit = this.page.getByRole('button', { name: 'Submit' }) fails with TypeError: Cannot read properties of undefined before the test body even runs. Fix: assign locators in the constructor. If you need the upgrade first, set "useDefineForClassFields": false in compilerOptions and move the fields over time.
2. Actions now wait for elements that are not inert. click(), check(), tap(), fill(), clear(), and selectOption() wait until the element is not inert, and time out if it stays inert. The nasty case is fill(): on 1.64.1, filling an input inside an inert form reported success and typed nothing. On 1.65 it times out and the call log says why. A test that starts failing here was probably never testing what you thought — check whether the form is inert because a dialog is still closing or a save is in flight.
3. Proxy credentials only go to the proxy. The username and password in the proxy launch option are no longer sent to websites; they answer only the proxy’s own challenge. A test that logged in to a site with Basic Auth through proxy credentials now gets a 401. Move site credentials to httpCredentials.
The other five:
| Change | What breaks | Fix |
|---|---|---|
| Cookie matching follows RFC 6265 | browserContext.cookies(url) no longer returns host-only cookies for subdomains or Path=/api cookies for /apiv2 | Ask for the right URL, or call cookies() with no URL and filter by name |
API params no longer replace the URL’s query | String/URLSearchParams params on a URL that already has ? are now appended, not replacing | Build the URL you mean explicitly |
| A sharded run with no tests now fails | npx playwright test --shard=1/2 --grep nomatch exits 1 instead of 0 | Add --pass-with-no-tests where empty is acceptable |
| Hidden text leaves accessible names and aria snapshots | Text inside a closed <details> or unassigned to a slot no longer counts; stored aria snapshots need updating | Regenerate with --update-snapshots, then read the diff instead of accepting it blindly |
| CSS imports in test files must exist | .css, .less, .scss imports now load as empty modules, so a missing file stops the run | Fix the path or delete the stale import |
Two smaller changes ride along: the blob reporter again empties its output folder before writing (give each job its own outputDir if you merge blob reports), and test locations in CommonJS output moved a few columns (only matters if a tool parses those locations). And one known issue to watch: navigator.platform emulation can cause Ctrl or Meta dispatching errors (#40009); the workaround while a patch ships is PLAYWRIGHT_NO_UA_PLATFORM=1.
The safe upgrade path: check Node.js is 22 or newer, fix page-object fields first (it breaks the most tests at once), run the full suite on a branch, and only then merge. A green run after these checks means something. A green run before them might mean nothing at all.
Why n8n builders are searching Playwright at all
The n8n connection is the MCP bridge. An n8n AI Agent node can reach a real browser through the MCP Client node pointed at @playwright/mcp — and since 1.59, that MCP server can attach to a browser your code already launched and logged in to, instead of starting a fresh, anonymous one. The pattern:
- A small Node script (or an n8n Execute Command node) launches Chromium and binds it:
const { endpoint } = await browser.bind('my-session', { workspaceDir: '/my/project' }). The browser stays alive, logged in, on the page your workflow cares about. - The n8n MCP Client node connects with
@playwright/mcp --endpoint=my-session. The agent gets browser tools — navigate, snapshot, click, fill — against that live session. - The agent’s work is observable:
playwright-cli showopens the dashboard, andpage.screencastcan record a video receipt of what the agent did, with chapters, for human review.
Figure 1 — The browser-bind machine. One live, logged-in browser, bound to a named session, handed to the agent through the MCP socket. Notice the lit bind lever, the session drum, and the screencast reel: the dashboard and the recording are the audit trail, not the permission model. Readout: 1.65 · 8 breaking changes · Chromium 158 · bind in 1.59.
That is the same shape as the approval and containment patterns this blog has covered for other agent tools: the agent acts through a scoped, observable channel, not through a shared master credential. Two guardrails carry over directly. First, keep the bind local: browser.bind() defaults to a named pipe; the host/port WebSocket option exists, but exposing it turns your logged-in browser into a network service — do not do that on a shared host. Second, put an approval gate in front of any browser action that spends money, submits a form, or changes state — the same Always Ask default as the DigitalOcean Cuttlefish approval pattern and the Whop PayBox approval flow. A browser the agent can drive is a browser the agent can misuse; the dashboard and the screencast receipt are your audit log, not your permission model.
For the MCP side of the wiring, see How to Connect an MCP Server to n8n and the MCP + n8n troubleshooting guide — including the queue-mode tool-call fix in Fix n8n MCP Tool Calls in Queue Mode. For the containment thinking behind all of this, the Replit MicroVM sandbox article and the n8n 2.43 guardrails release are the closest parallels on this blog.
Limits and honest caveats
I did not run this upgrade against a live suite for this article — the breaking changes and fixes above are Playwright’s own documented behavior, cross-checked against the November 8 breaking-changes guide, not a hands-on migration. The trend signal itself is soft: Google’s rising lists shifted between two loads minutes apart, and “playwright” also rises for ordinary test-automation reasons — the n8n pairing is the signal, not proof that every searcher is an n8n builder. Bun mode is explicitly limited (no watch mode, no --only-changed import following, no custom ESM loaders), so treat it as a fast path for suites that fit, not a default. And the 1.59 agent features are a month old by now; if you are only now discovering them through the 1.65 spike, the upgrade checklist above is the price of catching up in one jump.
Key Takeaways
- Playwright 1.65 shipped November 5, 2026 — and it breaks things. Eight breaking changes, led by page-object fields, inert-element waits, and proxy credentials. Upgrade on a branch, fix the page objects first.
- The agent browser is 1.59’s story, not 1.65’s.
browser.bind(),page.screencastreceipts, the CLI dashboard, and--debug=clilanded in 1.59; WebMCP landed in 1.64. Do not blame 1.65 for features it did not ship — or credit it for them. - The n8n pattern is bind, then MCP. Launch and log in once, bind the session, point the n8n MCP Client node at
@playwright/mcp --endpoint=my-session. The agent drives a live, logged-in browser instead of a fresh anonymous one. - Keep the bind local and gate the actions. Named pipe by default; do not expose the WebSocket option on a shared host. Approval before any state-changing browser action, with the dashboard and screencast receipt as your audit trail.
- Node.js 20 is done. 1.65 is untested on it; Docker images ship Node.js 26. Check the runtime before you check the tests.
FAQ
Is Playwright 1.65 safe to upgrade to?
Yes, on a branch, after the checklist: Node.js 22+, page-object fields moved into constructors, proxy credentials moved to httpCredentials, aria snapshots regenerated and reviewed. The breaking changes are all documented and fixable; none of them is a reason to stay on 1.64.
Does 1.65 add browser.bind() or the CLI debugger for agents?
No. Both shipped in Playwright 1.59, along with page.screencast and the playwright-cli show dashboard. Version 1.64 added WebMCP support. Several secondary write-ups attribute these to 1.65; Playwright’s official release notes say otherwise.
How do I connect Playwright to an n8n AI Agent?
Through MCP: bind a launched browser with browser.bind('my-session'), then point n8n’s MCP Client node at @playwright/mcp --endpoint=my-session. The agent gets browser tools against your live, logged-in session. Keep the bind local and add an approval step before state-changing actions.
What is the fastest way to run Playwright tests under Bun?
bunx --bun playwright test. It skips Babel and its compilation cache, but there is no watch mode, --only-changed does not follow imports, and console output does not reach the trace or UI mode. Try it on a branch before making it your default.
Where do I check whether a Playwright release broke my suite?
The official release notes at playwright.dev list every breaking change per version. For 1.65, the eight changes are: useDefineForClassFields page-object fields, inert-element action waits, RFC 6265 cookie matching, proxy credentials, hidden text in accessible names, params appending, CSS imports that must exist, and sharded runs with no tests failing.