MultiOn

Sweet spot: a developer who has identified one or two web tasks where there is no API, where 80% reliability is good enough, and where the cost of running their own browser-agent stack outweighs paying MultiOn per run. The hosted layer is the genuine value — anti-bot mitigation and login-state management are work you would otherwise do yourself. Failure modes. Agent reliability on long flows is the honest blocker — 6 sequential steps at 85% each is roughly 38% end-to-end success. Plan for retries, human handoff, or hard caps on chain length. Cost runaway on long tasks is the second issue — a 50-step task hitting GPT-4-class on every step adds up fast. The third is strategic: AGI-0 is clearly the company's consumer focus now, and a developer-facing Agent API can become a second-priority product overnight. What to pilot. Pick one web task where the alternative is "we built a brittle Playwright script and it breaks weekly." Run it through MultiOn for two weeks against real production targets. Track success rate, cost per run, and how often you intervene. If it beats your script on success rate or eliminates 80% of the maintenance, the math works. If success drops below 70% or per-run cost exceeds the manual labor cost, build it yourself with browser-use instead.

MultiOn (now operating as AGI Inc, parent company behind the AGI-0 mobile agent) shipped one of the first widely-used browser-control agent products: an AI that drives a real Chrome browser to complete tasks for you. Tell it "find me a flight to Tokyo under $800 next month" or "renew my Amazon Prime subscription," and the agent navigates, clicks, fills forms, and reports back. Available as both a Chrome extension for end-users and an Agent API for developers wiring browser-agent capability into their own products. The Agent API is the technically interesting layer. It exposes a hosted browser-control runtime via REST — you POST a task in natural language, get back a session you can stream events from, and receive the final result (or human-handoff request). Compared with running browser-use yourself on your own infrastructure, MultiOn handles the headless-browser fleet, anti-bot mitigation, session persistence, and login state for you. Versus the open-source browser-use library, MultiOn is the hosted SaaS layer — you trade local control and free-of-cost for someone else managing Playwright clusters and CAPTCHAs. Versus Anthropic's computer-use API, MultiOn is web-only (no desktop), but came earlier, sits at a higher level of abstraction (task-level rather than action-level), and prices for production use rather than developer experimentation. The company has been pivoting hard toward the consumer mobile-agent surface (AGI-0 on Android, with Visa/Qualcomm partnerships), so anyone evaluating the developer Agent API for production should confirm current SLA and roadmap directly. The technology is real and works; the question is whether the API is still the company's priority.