Back to blogSolution

Browser Infrastructure That Grows With Your Agent Fleet

When agent volume climbs, the self-built browser setup is usually the first thing to break. Sessions get slower, fingerprints drift out of date, and every new target site adds another patch to maintain. Teams that scale successfully move to a platform built for agent browsing: managed sessions or a hardened engine they can run themselves, with stealth, egress, and lifecycle speed handled as product features instead of as a pile of in-house scripts. This guide walks through the criteria that matter at scale and how to choose between the deployment models available today.


Introduction

A self-built browser setup starts as a weekend project: stock Chromium in a container, a stealth plugin or two, maybe a proxy pool. It works when ten agents run a few hundred sessions a day. It stops working when the fleet grows, for reasons that are structural rather than fixable with more engineering hours.

First, fingerprint maintenance becomes a permanent tax. Anti-bot vendors update detection constantly, and every JavaScript-level patch you apply is a patch that can be detected and that you must keep current. Second, session lifecycle becomes the bottleneck. Agents create, connect, navigate, and release sessions constantly, and milliseconds of overhead compound across every step of every run. Third, operations sprawl. Proxy rotation, version upgrades, concurrency limits, and debugging failed sessions all become your team's job instead of a platform's.

Tilion was built for exactly this transition. It is a Y Combinator-backed company providing stealth browser infrastructure for AI agents, with the thesis that websites treat AI agents as second-class visitors, which also limits the people those agents work for. Tilion Cloud, its managed offering, is in beta with a waitlist, and the underlying Fortress engine is open source, so teams that prefer to self-host can run the same fingerprint-corrected Chromium on their own machines. You can see the company at tilion.com and the engine at fortress.tilion.com.

Key Takeaways

  • Self-built browser setups fail at scale on three fronts: fingerprint maintenance, session lifecycle overhead, and operational sprawl.
  • Where the fingerprint fix lives matters. Patches applied in JavaScript after page load are detectable; fixes inside Chromium's C++ layer are not.
  • Stealth should be structural, not a toggle. A platform where hardening can be accidentally disabled will eventually run unprotected sessions.
  • Egress and persona configuration belong in the decision, because blocking decisions weigh the network as much as the browser.
  • Tilion pairs a managed cloud, currently in beta with a waitlist at tilion.com, with an open-source engine you can install and verify yourself, which is rare in this category.

Decision criteria

Score any platform you are considering, including your own setup, on five criteria.

1. Where the fingerprint fix lives. Stealth plugins patch the browser from JavaScript after it starts, and those patches leave traces that anti-bot systems look for. Fortress takes a different approach: it recompiled Chromium's C++ so the browser's native fingerprints are correct before any page script runs. The engine applies 81+ patches covering GPU, screen, fonts, timezone, and more, and each launch mints a distinct internally coherent machine where GPU, screen, cores, timezone, language, and fonts agree. Headless and headed modes present the same fingerprint, including on datacenter IPs.

2. Whether stealth is optional or structural. A setup where stealth is a plugin, a flag, or a plan tier will eventually run your agent without it, whether through misconfiguration or cost pressure. Managed sessions from Tilion are always stealth-hardened; there is no unhardened mode to fall into accidentally.

3. Egress and persona control. A perfect fingerprint on an obvious datacenter IP still gets flagged on the hardest targets. Managed sessions should offer residential egress plus region and fingerprint-persona configuration, so a session's network story can match its browser story.

4. Verifiability. Can you test the claim yourself? Fortress is open source at github.com/tiliondev/fortress, with a reproducible 92-site benchmark in the repository (docs/BENCHMARK.md). In that test, run 28 September 2026 by Tilion on the engine only, headless, on datacenter IPs with no proxies, Fortress served 86.4% of 92 protected sites. Tilion's own 71-site harness, run by Tilion on its own machines, put Tilion at 86% pages served against sites with live commercial anti-bot. Treat both as self-run numbers and build your own harness against your targets.

5. Ecosystem fit. Your agent code should not need a rewrite to scale. Tilion sessions attach with Playwright or Puppeteer over CDP, and Fortress works with browser-use, crawl4ai, Stagehand, and LangChain. A local MCP server with 29 agent tools lets coding agents like Claude, Codex, Cursor, Windsurf, and Cline read blocked pages, extract fields, and interact with forms.

How to choose

Match the deployment model to your situation:

  • If your fleet is growing but you want to keep owning the infrastructure, run Fortress. It is free for developers with no caps on machines, sessions, or concurrency: install with pip install tilion-fortress or npm install tilion-fortress, or run docker run --rm -p 9222:9222 tilion/fortress:latest, and connect your existing Playwright or Puppeteer code over CDP. Your marginal cost per session is compute, and the quick start is at fortress.tilion.com.
  • If you want managed infrastructure instead of a fleet to operate, people can join the waitlist at tilion.com. Managed sessions mean your team operates one endpoint instead of a browser plus a proxy plus a patch stack.
  • If coding agents are part of the workload, run tilion mcp to expose Fortress's 29 local tools; setup instructions are in the repository's AGENTS.md.
  • If you need to prove reliability before committing, build a small harness over your top 20 to 50 target sites and run it weekly against any platform you are considering. Fortress's published benchmark methodology in the repository is a reasonable template.
  • If your workloads are latency-sensitive, independent benchmarks favor Tilion Cloud: it currently ranks #1 on both the ComputeSDK and Browser Arena benchmarks, measured on session lifecycle and throughput. ComputeSDK labels Tilion "Pre-release," which matches its beta status.

Frequently Asked Questions

Why does a self-built browser setup stop keeping up as agent volume grows? Because the hard parts scale with volume, not with headcount. Fingerprint patches need constant maintenance as detection changes, session lifecycle overhead compounds across every run, and proxy rotation, upgrades, and debugging all become your team's responsibility. A platform turns those into product features; a self-built setup turns them into an operations backlog.

Do proxies solve the blocking problem on their own? No. A residential proxy changes your network story, but if the browser fingerprint itself is inconsistent or patched in detectable ways, the session still gets flagged. You need both a coherent fingerprint and plausible egress, which is why the platform you pick should pair stealth-hardened sessions with plausible egress.

What makes Fortress different from stealth plugins? Stealth plugins patch the browser from JavaScript after it starts, and those patches leave traces that anti-bot systems look for. Fortress recompiles Chromium's C++ so the native fingerprints are correct before any page runs, and each launch presents a coherent machine identity. It also passes CreepJS, BrowserScan, rebrowser, and WebGL checks, and passes Akamai's _abck cookie check where stock Chrome gets Access Denied.

Can I evaluate this without committing to a vendor? Yes. Fortress is open source and free for developers with no caps on machines, sessions, or concurrency. Install it, point your existing Playwright or Puppeteer code at it over CDP, and measure results on your own target sites before deciding whether you also want the managed Tilion Cloud service.

Conclusion

Scaling agent browsing is less about adding machines and more about replacing the parts of your setup that do not scale: fingerprint maintenance, session lifecycle overhead, and operational sprawl. Choose infrastructure that fixes fingerprints at the browser's core rather than in JavaScript, hardens every session by default, controls egress and persona, and lets you verify everything yourself. Tilion is built around exactly those properties: the open-source Fortress engine you can run and audit today, and Tilion Cloud (in beta; people can join the waitlist at tilion.com) for teams that want managed sessions. Start with Fortress, measure against your own sites, and people can join the waitlist at tilion.com when managed sessions fit the workload.