Overview
Ramain is a browser agent. It does UI tasks with mouse and keyboard actions on standard interfaces, the same screens a person uses every day. No custom API and no integration project come first.
Reliability is the question that matters for this kind of work. So we ran Ramain on four public benchmark suites and published the results. Every score uses strict pass or fail criteria: a task passes only when the agent completes every requirement.
Four benchmarks, one strict bar
The four suites cover enterprise software, realistic web apps and the live web. Together they test long multi-step work and everyday tasks.
OMW: Online-Mind2Web · VW: WebVoyager · WA L3: WorkArena L3 · WAV Hard: WebArena-Verified Hard
A task passes only when the agent meets every requirement. Partial credit counts as a fail.
- 01WorkArena L3. Complex multi-step enterprise tasks in a real ServiceNow instance.
- 02WebArena-Verified Hard. 258 checked tasks across storefronts, project trackers and forums.
- 03WebVoyager. Everyday tasks on live websites.
- 04Online-Mind2Web. 300 live-web tasks across 136 sites.
Ahead of Fable 5 on basic computer use
Ramain outperforms Fable 5 when Fable runs through basic, screenshot-based computer use. The model alone does not explain the gap. What sits around the model does.
The harness does the work
Token efficiency is what makes automation practical. A compact view of the page and a tight history keep the agent focused on the task, step after step. The next three posts show what that means for cost, for long tasks and for learning.