Browser Run now runs on Cloudflare Containers, which lifts its usage ceilings, cuts Quick Action latency and gives the team its own browser image to upgrade. The changes are already live; existing workers need no redeployment or code changes.
The headline numbers: 60 browsers can be created per minute through the Workers binding, and 120 can run at once — 4x the previous ceiling. Quick Action response times have fallen by more than 50%.
Why the old shared setup had to go
Browser Run lets developers drive headless browser instances on Cloudflare's network, for tasks such as end-to-end testing, inspecting suspicious URLs, rendering PDFs, taking screenshots, extracting content and, increasingly, agent workloads that interact with the web.
Those browsers previously shared infrastructure with Browser Isolation (BISO). The two are technically similar, but the arrangement hurt Browser Run in several ways: BISO's larger container images slowed startup, BISO browsers were not distributed optimally by region (costing resiliency and latency), and BISO's long, steady sessions conflicted with Browser Run's short, spiky traffic, producing scaling bottlenecks and availability delays.
With Durable Object-enabled Containers in open beta, the team began a tentative migration that ended up benefiting both platforms.
How the cutover was staged
A Worker inserted into the incoming request path let the team serve some browsers from Containers while others still came from BISO. Running both side by side during development made it possible to compare performance and flush out implementation bugs.
The rollout then progressed in stages: all Quick Actions endpoints first, then the Workers browser binding for free accounts, then pay-as-you-go accounts, and finally the remaining contract customers once stability was validated.
Keeping Durable Objects and containers close
The migration introduced problems beyond the early-stage platform's thin documentation, limited observability and scarce colleagues in overlapping timezones. Feedback as an internal "Customer Zero" fed fixes back into the platform itself.
The architecture also had to adjust. A DO-enabled Container creates its Durable Object as near as possible to the incoming request, but the container it connects to may spin up on the other side of the world. That is tolerable for a single message like "start my app," yet a screenshot request rides a WebSocket and exchanges dozens of messages — milliseconds spent crossing the globe accumulate.
The fix was regional pools of pre-warmed DO-backed browser containers, which cap the maximum distance, and therefore latency, between a DO and its container. Each request gets the closest DO-container pair within the region, keeping both hops short. The design adds moving parts, but it works provided capacity can be reallocated as demand shifts — which put the pool state on the critical path.
Trading KV for D1 and Queues
Container state originally lived in Workers KV, which was acceptable at first but whose eventual consistency — cache TTL was recently lowered to a 30-second minimum — turned into a bottleneck as AI agent builders pushed request volumes past existing capacity. A container could read as "available," but be claimed before the request routed to it 30 seconds later, causing race conditions and over-allocation that throttled scaling during demand spikes.
State moved to D1, whose transactional nature fits a model where an assigned browser belongs exclusively to one user. SQLite transactions make assignment atomic, so two requests cannot claim the same browser. A simplified form of the acquisition query:
WITH candidate_pool AS (
-- candidate pool logic to pick based on latency and other rules
)
UPDATE containers
SET status = 'picked'
WHERE sessionId IN (
SELECT sessionId
FROM candidate_pool
ORDER BY RANDOM()
LIMIT ?5
)
RETURNING data
D1 shards are kept per location. With several thousand containers each writing state every 5 seconds, individual writes would overload the database: at 1ms per write and one row per write, roughly 5,000 containers is the breaking point.
Batching removes the ceiling, since a batch takes little longer than a single write. With 100-row batches, a location can now update up to 500,000 containers, and P95 batch write latency sits at 0.1ms.
Every 5 seconds each container computes its state and enqueues it on its location's queue. A Worker consumer is configured with a batch size of 100 and a 1-second batch timeout:
{
...
"queues": {
"consumers": [
{
"queue": "production-core-containers-queue-weur",
"max_batch_size": 100,
"max_batch_timeout": 1,
"max_retries": 1,
},
...
]
...
}
}
Lag stays comfortably under 2 seconds. If a backlog does produce stale state, each region falls back to a designated backup region until its primary queue catches up.
Faster Quick Actions
Dedicated infrastructure meant the container image could be changed without bloating or disrupting other products such as BISO, opening the door to optimizing screenshots and content extraction. The old flow opened a WebSocket to the remote browser and issued instructions one at a time — load the page, navigate, wait, capture — with each step gated on the one before it. Now all parameters go in a single HTTP request to the container, and the whole flow executes internally with no back-and-forth between Worker and browser.
Average quick-action response time has dropped sharply, with users waiting less for a browser to be ready and getting faster processing of their DevTools Protocol messages.

Ownership pays off in browser features
Sharing BISO's containers also meant upgrading Chrome required coordinating across teams with their own roadmaps and priorities. Running its own image lets Browser Run upgrade faster and set version and flags without side effects elsewhere. WebGL, a much-requested feature, is now available for browser-based rendering, alongside WebMCP (Model Context Protocol for the web), which enables new agentic interaction patterns. Solving state management at this scale has also freed time for development such as the recently launched /crawl endpoint.
Browser Run is available on all Workers plans. The quick start guide and Quick Actions documentation cover the REST API, including the /crawl endpoint for following links across a site. The Agents SDK ships with built-in Browser Run support.



