Hermes Agent Cuts 12 Browser Tools Down to One

Twelve tools become one

Nous Research announced it has applied a new "browser use mode" to its agent, Hermes. Previously, the agent relied on twelve separate tools to handle browser actions like clicking, scrolling, and typing. These have now been replaced with a single tool powered by browser_use's CLI 3.0. The company explained that "instead of calling a tool for every click, the agent writes a script." Internal testing reportedly showed token usage dropping by 48-66% with no loss in accuracy.

Why tool calls were so expensive

When an AI agent operates a web browser, it normally has to send the LLM a separate request for every single action — "click this button," "type into this field," and so on. This method is called a "tool call," and each request requires re-feeding the model the full description (schema) of all twelve available tools. Since this entire manual gets attached every single time, even for one click, the wasted tokens — the smallest unit an LLM uses to process text, and the basis for pricing and speed — pile up quickly as conversations get longer.

Nous Research took a different approach: having the agent write out multiple actions at once as a code script and execute them together. This eliminates the need to resend the twelve-tool manual for every click. This "represent actions as code" approach echoes a direction several researchers in agent design circles have recently been exploring. Nous Research is known as the developer behind the open-source LLM family Hermes, and this latest announcement appears to be another experiment aimed at the open-source ecosystem. browser_use is a browser automation framework that lets AI agents click, type, and navigate real websites, and its CLI (command-line interface) 3.0 now serves as the engine driving Hermes's new mode.

So what changes

AI agents that operate browsers have become an area where both cloud providers and startups are racing to release products. The persistent problem has been cost — the more steps a web page task requires, the more tokens get consumed, making it too expensive to deploy in real services. If this improvement holds up, it means the same task could be handled with less than half the tokens, which could help lower the operating costs of services that run agents continuously. However, this is based on Nous Research's own internal testing, and it has not been confirmed whether the results have been verified by independent benchmarks.

Claude-based AI assistant hacks gym system to cut in line for booking