CAPTCHA walls in Playwright agent pipelines are usually self-inflicted retry bursts

Pacing, not the session cap, is what the far end actually sees. In the reference pool, every `acquire()` reads a per-domain last-hit timestamp and sleeps out the remainder of 2 seconds before handing back a session ^[8]. Retries pass through the same gate, because `execute()` releases the session on failure and calls `acquire()` again for the next attempt ^[8]. A failure therefore cannot produce an immediate second request, and ten simultaneous failures cannot produce ten simultaneous retries; they line up behind one timestamp.

The arithmetic follows from that. One admission every 2 seconds is 30 tasks per minute per domain, whatever `maxActive` is set to ^[1]. Five slots only begin to bind when a task holds one for longer than 5 x 2 = 10 seconds ^[2]. On single-domain work with page steps shorter than that, tuning the concurrency number does nothing and tuning the interval does everything. Spread the same work across twenty domains and it inverts: pacing is tracked per domain, so the cap is the only thing holding total context count down ^[8].

The other figure worth budgeting is the worst-case task. A 30-second timeout, three attempts, and backoff of 2 then 4 seconds put 96 seconds between first attempt and thrown error ^[3]. Queue depth and any caller-side timeout have to sit above that number.

Then there is the listing itself. `acquire()` throws `Queue full` when `this.queue.length` reaches the `maxQueued` value of 10, and `this.queue` is initialised to an empty array that nothing in the published code ever pushes to [c8b]. The cap never fires. Overflow instead becomes an unbounded set of callers each spinning on a 100ms `setTimeout` while waiting for a slot ^[8]. The prose sells overflow queuing; the object does admission control. The Redis variant does `rpush` and does enforce its limit of 50 pending tasks ^[9], which reads like the in-process class was trimmed for the article rather than run in anger.

What the dev.to post establishes is a failure mode described in enough detail to recognise in your own logs, and a remedy specific down to the millisecond. What it does not establish is the size of the win: there is no CAPTCHA rate, no throughput figure and no before-and-after behind the claim that 5 active sessions plus 5 queued beats 50 tabs ^[6][11]. That is survivable, because the claim is cheap to falsify. A semaphore and a timestamp map cost an afternoon, and if the CAPTCHA rate does not move afterwards, you have learned something about the target's defenses that no proxy vendor was going to tell you ^[2].