Agent Browser Fixes the Queue Before the Model • The Weighted Average

Agent Browser 0.33.2 cut a click after a deliberately bursty 300-mouse-event sweep from 2,471 milliseconds late to 6 milliseconds. For a transparent operation watching 100 sessions at 10 frames per second, its smallest measured image profile also removes 28.5 TB of monthly traffic—evidence that browser-agent latency is often a queue problem before it is a model problem.

Three queues, one misleading latency number

The official 0.33.2 release changes input dispatch, application-frame selection, and per-client delivery. Before the fix, the WebSocket reader awaited a Chrome DevTools Protocol response for every mouse move, serializing a subsequent click behind 300 acknowledgements. PR #1627 reports the 2,471-to-6-millisecond result, a 2,465 ms or 99.76% reduction, while preserving click-pair, drag, and typed-character order.

That scope is the story. It is a maintainer-run stress sweep, not universal model-to-action latency. Selector-driven clicks with no mouse backlog may barely move. The same test found ack-paced video holding 54–56 fps from 30 to 250 input messages per second, versus 3.7 fps at 200/s before, showing how serialized control traffic had starved an otherwise unrelated frame path.

The new architecture separates three forms of backpressure. PR #1594 explains that Chrome frames are acknowledged upstream immediately, while a latest-wins watch channel retains only the newest application frame. Each external WebSocket splits reading input from writing frames, and optional renderer acknowledgements allow only one frame in flight. Chrome’s protocol defines the screencast and acknowledgement primitives; Agent Browser composes them into a freshness policy.

Latest-wins alone cannot erase bytes already buffered into a socket. In a constrained-link test at 400 KB/s with an eight-second client stall, push mode later drew 436 stale frames and 7.7 MB; ack pacing drew one frame and 0.02 MB. A proxy that acknowledges on receipt rather than after render can reintroduce a downstream queue, because the browser exposes WebSocket bufferedAmount as queued bytes. The ack must terminate at the viewer that establishes freshness.

Encoding is the larger economic lever. PR #1626 measured 54 KB per frame at quality 80 and 1280×720, 25 KB at quality 20 and the same dimensions, and 9 KB at quality 20 and 640×360. The smallest profile cuts bytes 83.33%; quality alone cuts 53.7%. Those are one busy page’s measurements, not universal JPEG ratios.

The smallest stream profile cuts frame bytes 83%

Measured JPEG kilobytes per frame on one busy page

Vercel Labs agent-browser PR #1626 · Aug 2026

Now state the scenario: 100 concurrent observed sessions, 10 fps, eight hours per day, 22 days per month. That is 17,600 browser-hours. At 54 KB, traffic equals 54 × 10 × 3,600 × 17,600, or 34,214.4 GB. At 9 KB, it is 5,702.4 GB. The difference is 28,512 GB, or 28.5 TB per month. At a transparent assumed $0.09/GB marginal egress rate, the saving would be $2,566—not a quoted Agent Browser fee.

This turns today’s Meta billing lead inside out: both show that a tiny technical unit becomes material only after workload volume and semantics are specified. It also extends the Prentis computer-use lab’s cost-per-successful-task argument from model selection into transport design.

Upgrade stream users, benchmark the pixels

Teams using live WebSocket video for human takeover, streamed QA, debugging, or cursor-heavy computer use should canary 0.33.2 now. The repository describes streaming as an optional surface of a native Rust CLI controlling Chrome, so accessibility-snapshot or direct-command users should take the bug fixes without expecting a 2.5-second inference improvement.

A practical canary uses pacing=ack&maxFps=10, sends the ack only after render, and compares the three published quality profiles against task success. The implementation cost is modest for the bundled dashboard but nontrivial for a custom consumer: it must handle sequence numbers, deliberately dropped intermediate frames, capture timestamps, reconnect behavior, and staleness telemetry. Latest-wins video is a viewer, not an audit recording.

The browser-hour impact is smaller than the traffic impact in ordinary work. Under an intentionally aggressive scenario—100 agents, 200 clicks a day, 22 days, and every click following that same 300-event sweep—440,000 clicks save 301.28 browser-hours. Browserbase lists startup overage at $0.10 per browser-hour, making direct occupied-session savings only about $30, although reclaimed capacity equals 1.71% of the scenario’s 17,600 hours. Do not turn a stress calculation into a normal-click forecast.

Quality can break the bandwidth verdict. A 640×360, quality-20 stream may impair OCR, tiny-target grounding, or human diagnosis. Ack pacing can also cap frame rate near one frame per transfer-plus-render round trip on high-latency links. No-await input dispatch could move congestion into Chrome or the CDP socket under extreme floods. The release evidence lacks machine, page, Chrome version, RTT, and p95/p99 distributions.

The operator test therefore needs four clocks: model think time, command-queue time, CDP dispatch-to-effect, and capture-to-render age. Track task success, bufferedAmount, browser-hours, and egress separately. Reverse the low-resolution recommendation if success falls by more than one percentage point; narrow the latency claim if production replays across several RTT bands improve p95 click-to-DOM effect by less than 100 ms.

The broader engineering lesson matches the archive’s coding-agent memory analysis: preserve the state that matters and discard obsolete work deliberately. For video, freshness beats completeness. For control events, order beats synchronous acknowledgement. For audit, a separate complete recording remains necessary.

Upgrade if the stream is on the critical path. Then benchmark pixels before celebrating bytes. Version 0.33.2 does not make a model think faster; it stops yesterday’s frame and the previous mouse move from holding tomorrow’s click hostage.