Concurrent coding agents were contending for one Chrome profile. I built a TypeScript CLI and session library that gives each worker its own browser, waits for readiness, and cleans up the session when work ends.
A worker requests a browser, navigates to a page, reads its title, and captures a screenshot. Each worker has a distinct port and temporary profile, so another worker’s navigation cannot replace its page.
My code manages allocation, readiness, attachment, and teardown. Playwright drives the page through Chrome DevTools Protocol. The agent or calling script chooses the task and interprets the result.
Separate state for each task, with an explicit end to its lifetime.
Port reservation and its placeholder record are written inside the allocation lock. Readiness polling happens outside it, so one slow browser does not hold up every other startup. The default readiness limit is 15 seconds; a timeout removes the browser and its temporary state.
The session wrapper releases the browser in a finally block. A live integration test deliberately throws inside a task, checks that the error reaches the caller, and verifies that no live instance or orphan remains. The screenshot demo also allows one retry of its initial navigation.
Disconnecting a CDP client can leave Chrome running. Release finds the processes associated with that session’s profile, stops them, and verifies they are gone. A diagnostic command reports abandoned profiles and processes within the tool’s own storage area.
September 29, 2026 UTC · Linux CI · Headless Chrome
Four workers visited different public pages and saved screenshots. The integration checks covered navigation and release, distinct concurrent ports, bulk cleanup in an isolated test environment, and cleanup after a thrown task. The same revision passed 93 unit tests across 12 suites.
This is one recorded run, with network-dependent timing and no separately measured sequential baseline. It verifies browser lifecycle behavior; it does not measure an LLM’s success rate on arbitrary computer tasks or exercise the macOS-specific launch path.
Download the measurement summary (JSON)