TL;DR
- Browser infrastructure runs isolated browser sessions and gives agents control over pages. An agent runtime manages the task and decides which actions to take. Some products span both layers.
- Compare providers against the requirements your workload actually imposes, especially session isolation, persistent authentication, human intervention, observability, security, browser control, and cost.
- Anchor fits workloads that need stealth options, high-scale session launches, or authenticated account workflows. Anchor reports batch creation of up to 5,000 sessions and Enterprise launch throughput of 100,000 or more sessions per minute. Neither figure describes active concurrency.
- Evaluate Browserbase and Hyperbrowser when you want browser hosting alongside agent tooling. Compare Kernel, Steel, and Browser Use against your preferred level of control over the agent and browser layers.
- No provider wins every deployment. Test your likely session volume, authentication steps, and intervention needs before choosing.
Browser execution layer vs. agent runtime
Browser infrastructure supplies the environment where an agent uses a browser. It creates isolated sessions and gives you ways to retain authentication, inspect activity, and control the browser. The agent-side software decides which action to take next and sends that action to the browser.
Those responsibilities have a finer split in production. Restate distinguishes the framework that supplies building blocks, the harness that runs the reasoning loop, and the runtime that manages state, recovery, and operations. A browser sandbox handles execution separately, although a product can combine these roles. AWS AgentCore illustrates session isolation with a dedicated microVM for each user session. That example describes an architectural pattern, not a capability claim about any provider in this comparison.
Several providers offer products on both sides of the boundary. Browserbase Agents, Hyperbrowser’s HyperAgent, and Browser Use’s hosted agents warrant separate evaluation from their browser execution services. Traditional RPA can use browser infrastructure as well, but its predefined workflows differ from an agent loop that chooses actions as it runs.
Comparison: which provider fits which production need
Browser infrastructure runs isolated browser sessions that an agent can control. An agent runtime decides which actions to take. Browserbase, Hyperbrowser, and Browser Use offer products at both layers, so compare the browser service separately from any hosted agent.
The concurrency figures in this table are provider-published plan figures, not independently tested capacity. Batch size, active concurrency, and launch throughput measure different things.
- Anchor best for high-scale or authenticated workflows needing managed browsers and stealth options.
- Browserbase best for browser sessions paired with Stagehand or hosted agent tooling.
- Kernel best for a separately built agent runtime needing isolated browsers and managed auth.
- Steel best for self-hosted browser execution with familiar automation tools.
- Hyperbrowser best for cloud sessions with an optional packaged agent.
- Browser Use best for low-cost browser trials or a separately chosen open-source agent library.
Anchor Browser: stealth, scale, and authenticated enterprise workflows
Anchor Browser fits production workloads where bot-detection risk, browser volume, or authentication lifecycles shape the infrastructure choice. Anchor Chromium combines humanized browser behavior with proxy and network options to reduce detection risk during interactive sessions. Web Unlocker serves high-volume page retrieval, while browser sessions support workflows that need clicks, form submissions, or an authenticated user. These stealth capabilities do not guarantee bypass of any site’s protections.
Anchor reports three distinct capacity measures. You can create up to 5,000 browser sessions in a batch. Its Enterprise program supports launch throughput of 100,000 or more sessions per minute, which measures how quickly sessions start, not how many run at once. Plan concurrency measures simultaneous browsers. The Free plan allows up to five, while Enterprise includes 500 or more concurrent browsers. You should size each measure against the workload it describes.
For account-connected workflows, OmniConnect manages authentication handoffs, including one-time passcodes, MFA, re-authentication, and session recovery. Anchor can pause an agent for human intervention and resume after a response. Web Action Cache lets you reuse deterministic browser steps when a workflow does not need the agent to decide each action, reducing repeated model calls.
Browserbase: agent-layer tooling with managed browser sessions
Browserbase fits teams that want managed browser sessions and agent tooling from the same provider. Its browser sessions work with Playwright, Puppeteer, and Selenium, while Stagehand supplies agent automation primitives. Browserbase Agents can run tasks, and Functions can host TypeScript code. You can use the sessions with your own agent runtime or use Browserbase’s higher-level tools to decide and execute actions.
Browserbase supports authenticated work across sessions through contexts, which retain cookies and application data. For troubleshooting, its observability tools provide logs and metrics, while live view lets a person inspect an active session. Browserbase also documents session replay and a separate option to pause an Agent run for input.
Browserbase’s published pricing sets concurrency at 3 browsers on Free, 25 on Developer, and 100 on Startup. Its custom Scale tier starts at 250 concurrent browsers. Those figures describe simultaneous capacity, not session launch throughput. Browser hours beyond included allowances and optional services can raise the bill. For enterprise deployment, Browserbase lists private-cloud options and VPC connectivity, but does not offer on-premises deployment.
Kernel: fast session creation with managed authentication
Kernel fits engineers building their own agent loop who want managed browsers with credential controls. Kernel’s documentation assigns action decisions to the agent framework while Kernel supplies browser infrastructure. Kernel reports approximately 30 milliseconds for median browser creation and says each Chromium browser runs in its own isolated VM. The speed figure is a vendor claim, not an independent benchmark.
For authenticated workflows, managed authentication supports connection recovery, and saved profiles retain browser state. Kernel’s vault filling enters credentials without exposing their values to your app or model. You can inspect active sessions through live view, review replays, and send telemetry to your monitoring tools.
Kernel’s listed plans allow 5 concurrent browsers on Free, 10 on Hobbyist, and 150 on Start-up. Hobbyist costs $30 per month and Start-up costs $200 per month, both plus usage. Those concurrency limits describe browsers running at once, not creation speed. Kernel bills usage per GB-second, so a browser-hour comparison depends on the resources each session consumes.
Steel: open-source browser infrastructure for agent workloads
Steel fits you if you want an open-source browser execution layer that you can self-host or connect to your own agent runtime. Steel provides Docker and clustered deployment options, while Playwright, Puppeteer, and Selenium integrations let your agent control browser sessions. Your agent still decides which actions to take.
For authenticated work, Steel supports persistent profiles and reusable authentication context. You can view a live session and take over when an agent needs help. Steel’s agent traces sync actions with session video, which helps you inspect what happened during a run.
Steel’s published pricing separates browser time from proxy traffic and CAPTCHA solves. The Launch plan lists $0.10 per browser hour and 10 concurrent sessions. Scale lists $0.08 per browser hour, a $250 monthly plan fee, and 100 concurrent sessions. Those browser-hour rates exclude proxy and CAPTCHA charges. Steel lists its stealth browser under Enterprise, alongside custom pricing and 1,000 or more concurrent sessions.
Hyperbrowser: cloud sessions with an optional agent layer
Hyperbrowser fits you if you want managed cloud browser sessions with the option to add a packaged agent. You can control sessions through Playwright while your own runtime decides which actions to take. Alternatively, HyperAgent supplies an adjacent agent layer.
For authenticated workflows, profiles preserve browser state across sessions, and stored credentials can fill login fields without exposing credential values to your code or the model. You can view and interact with a live session when an agent needs help, then use recordings to inspect what happened afterward.
Hyperbrowser’s pricing page lists one concurrent browser on Free, 25 on Startup at $30 per month, 100 on Scale at $100 per month, and 1,000 or more on custom Enterprise plans. Those figures describe active concurrency, not session-launch speed. Usage draws credits across different charges. Hyperbrowser lists base browser time at 100 credits per hour, proxy traffic at $10 per GB, and HyperAgent at 20 credits per step, so your cost estimate should account for the services each task uses.
Browser Use: low-cost browser infrastructure and hosted agents
Browser Use fits cost-sensitive prototyping when you want to choose between running your own agent and paying for hosted runs. Its open-source Python library supplies an agent runtime that can connect to browsers outside Browser Use Cloud. Separately, Browser Use Cloud offers browser sessions you control through Playwright or Puppeteer and hosted agents that run tasks for you. The library, cloud browsers, and hosted agents carry different responsibilities and costs.
Browser Use Cloud covers several needs beyond a temporary browser session. Profiles preserve cookies and browser state for repeat work, while hosted agents support live human takeover and continuation. You can inspect ordered run events, and you can enable session recording, which starts off by default.
Browser Use lists pay-as-you-go prices of $0.02 per browser-hour and $5 per GB of residential proxy traffic, with browser time rounded up to a one-minute minimum. Hosted agents add model cost plus 20%, alongside browser time and traffic. The pricing page lists 10 starting concurrent sessions and places 50 at $200 in lifetime spend, but the documentation index lists a $100 threshold for 50. Verify your account’s concurrency threshold and expected total cost with Browser Use.
Matching providers to deployment scenarios
For high-volume page extraction where bot detection affects completion rates, shortlist Anchor. Its Web Unlocker supports page retrieval, while its managed browsers support interactive work. If you need large browser fleets, assess batch creation, Enterprise launch throughput, and active concurrency as separate requirements.
For a SaaS product that connects customer accounts, Anchor fits workflows involving MFA handoffs, re-authentication, and session recovery through OmniConnect. If you plan to build your own agent loop and want managed authentication alongside isolated browsers, Kernel offers a browser execution layer that your agent framework can control.
For a self-hosted browser execution layer, Steel is the clearer starting point because it documents Docker deployment. You can connect your own agent to its sessions rather than adopt a packaged task runner.
For cost-capped prototyping, Browser Use Cloud lists browser time at $0.02 per hour, with a one-minute minimum. Budget separately for proxies and hosted agent runs. If you want packaged agent tooling with managed sessions, compare Browserbase’s Agents and Stagehand with Hyperbrowser’s HyperAgent. Browserbase and Hyperbrowser both offer browser sessions alongside agent-layer options, so you can choose how much of the task-running loop to build yourself.
FAQs
Does browser infrastructure replace my agent framework?
No. Browser infrastructure runs browser sessions and provides controls such as authentication and live view. Your agent framework still chooses what to do next, although some providers offer both browser infrastructure and hosted agents.
Can any provider guarantee bypassing bot detection?
No provider can guarantee access through a website’s protections. Anchor’s stealth capabilities can reduce detection risk, but the website still controls access.
How do batch creation, concurrency, and launch throughput differ?
Batch size counts sessions created in one batch. Concurrency counts browsers active at once, while launch throughput counts sessions started during a period. Anchor reports batch creation of up to 5,000 sessions and Enterprise launch throughput of 100,000 or more sessions per minute. Your plan’s concurrency limit determines how many browsers you can run simultaneously.
Choosing your browser infrastructure provider
Choose browser infrastructure by the requirement that constrains your workload, whether that is session isolation, persistent authentication, scale, stealth, or cost. No provider ranks first for every combination of those needs.
Test a representative workflow at your expected concurrency, including login handoffs and session recovery where they apply. Compare documented capacity and pricing with the completion rate and total cost you observe. For stealth-sensitive work, test against the sites your agents use rather than treating anti-detection claims as a guarantee.
