Keep exploring · more guides
Hermes AI just changed browser agents.
Nous Research dropped an update a few hours ago that takes twelve separate browser tools and replaces all of them with one.
In their own tests, it cut token use by 48 to 66 percent.
Same accuracy. Same results. Roughly half the cost.
By the end, you'll know exactly how it works, how to turn it on, and the one setup where it doesn't work at all.
One to navigate. One to click. One to type. One to scroll. And eight more.
The real twelve — these are the actual tool names from the Hermes docs.
Even to click one button, the AI hauled the instruction manual for every tool. Every single time.
Every single action was its own round trip back to the model.
What you're watching: the classic twelve-tool loop — one tool call per action, a full stop after every step. Recreated at reading pace from the real Hermes tool names; this is exactly the loop the update kills.
"Can I walk to aisle three?" Yes. "Can I pick up the milk?" Yes. You'd lose your mind — and burn a fortune in calls.
Instead of twelve tools, the agent gets one — and writes one script that handles the whole job.
This is the official drop — twelve browser tools replaced by a single one, driven by Browser Use's CLI 3.0. The agent writes a script instead of firing a tool call per click, and their own tests measured a 48–66% token cut with no accuracy drop.
What you're watching: a real Hermes session on this machine — one plain-English ask, the agent drives Browser Use CLI 3.0, and live Hacker News headlines come back. Real commands, real output, replayed at reading pace.
Understand this and you understand where all AI agents are heading.
Twelve instruction manuals shrink to one. The AI stops paying to carry what it isn't using.
The AI writes the full set of instructions up front — like your friend writing the whole shopping list.
Navigate, click, type, scrape, done. No phone calls home between steps.
With Opus 4.8 driving the browser, tasks that burned 50,000–64,000 tokens dropped into the low 20,000s.
Nous Research's own test numbers, posted this morning — same accuracy, roughly a third to half the tokens.
The tool manual shrank from a whole stack to one — and the agent now drives the browser with code, so the whole conversation gets shorter.
Cheaper agents, longer runs, bigger jobs, more runs per day — the token cost was the ceiling, and Nous just raised it.
Hermes is wired directly into our Agent Operating System — the same system where you plug in your Claude, your OpenClaw, your Free Claude Code, all sharing one memory.
No — that's the biggest myth about it. The everyday 90% runs on free local models and free APIs, and for the frontier work it drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI.
It's a layer on top of what you already own, not a new meter. And this exact update just cut the browser half of the bill again — plus the Boardroom has full token-efficiency tutorials inside.
If the Browser Use CLI can run on your machine, this mode is now the default — Hermes checks, and quietly falls back to the old twelve tools if it can't.
One config line to force it on. That's the whole setup — and if you'd rather keep the old tools, it's one command the other way.
What you're watching: the real update run on this machine — hermes update pulls the release, the new browser_exec tool lands, and the config line locks it on. Real commands, real output.
No. The new mode drives your own local browser for free — the Browser Use CLI is open source.
Their cloud browsers are an optional extra, and every other setup (Browserbase, your local Chrome, the built-in browser) keeps working with its existing credentials.
Wrong: "The agent writes Python scripts now — that means I need to be a coder to use this."
Right: The agent writes the code, runs it, and reports back. You type a plain sentence — "go to this website and fill in the contact form." If you can write a text message, you can run this.
Local Chrome, Browserbase, Browser Use cloud — the new mode sits on top and drives whichever browser you've already picked.
The privacy-focused local browser has no CDP endpoint for the new harness to attach to — Camofox setups automatically keep the old twelve tools. Nothing's broken; it's just not supported there yet.
Check if you're on Camofox — that's almost certainly why. And one honest note: because this mode runs real code on your machine, Hermes only offers it in sessions with terminal access. Locked-down setups (like a messaging-only surface) keep the standard tools — a safety choice, and the right one.
Hybrid routing: normal websites use your cloud browser, but anything on your own machine or home network gets a local browser automatically — the cloud provider never even sees it.
A shop owner's supplier-price watch every morning. A freelancer's hour of directory scraping — done while you're on a client call.
What you're watching: the Browser Use CLI 3.0 for real on this machine — it opens a live site, reads the page state with numbered elements the agent can act on, and closes clean. This is the exact harness Hermes now drives.
Wrong: "AI agents are too expensive to run every day. That's for big companies with big budgets."
Right: That was fair last week. But when a task drops from 50,000–64,000 tokens to the low 20,000s overnight, the math on "run this every single day" changes completely. Jobs that didn't make sense last Monday make sense now.
Wrong: "Browser agents look good in a video but they're not reliable enough for real work."
Right: Nous measured no accuracy drop — and the whole approach got simpler. One tool instead of twelve means fewer round trips, and fewer chances for the agent to lose the thread. Still check its work. Still start small. But "this stuff never works" gets more outdated every month.
Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries. Real businesses, real numbers, in their own words.
Read the 158-page wins doc →"A dozen schemas per click was never a capability problem, it was a token tax nobody totaled up until now."
— @tabor_bachelor, replying to the Nous Research announcement, Aug 11, 2026
Update it and check whether Browser Use mode is active — if the CLI is available, it probably already is. Run one small browser task and watch how much faster it feels.
Genuinely good week to start. It's free, it's open source from Nous Research, and browser automation just became one of its strongest features. Give it one boring job you do every week.
Pick a real, repeating, annoying task from your business — not a test, not a toy. Hand it to the agent. The moment you watch it complete a real job, something clicks about what's now possible.
This update handed everyone a browser agent that costs roughly half as much to run as it did yesterday. The question is whether you'll actually use it — and inside the AI Profit Boardroom, we've built the setup path for exactly this.
Twelve tools became one.
Token costs dropped by more than half.
Accuracy stayed the same.
That's what progress in AI looks like right now — not always a flashy new model, but the running costs of useful work quietly falling through the floor.
The gap between people who use this stuff and people who watch videos about it grows a little wider every week.
Pick one boring job this week. Hand it to an agent. Start there.
