Transcription
On May 10th, the most used AI agent on the planet changed. Not OpenAI, not Google. An open-source agent called Hermes, built by a research lab most people have never heard of, just processed over 560 billion tokens in a single day and knocked OpenClaw off the top spot it had held since it launched.
If you run agents in your business or you're about to, this flip tells you something important about which ones are actually safe to build on. Quick context on me. I've been building AI systems for founders for years and I watched the agent leaderboards the way some people watch the stock market. This particular move caught my eye. Here's why.
Open on the raw number. 568 billion tokens in one day versus OpenClaw's 151 billion. The timeline is the spectacle. Hermes launched February 25th. It hit number one in roughly 90 days. 140,000 plus GitHub stars, nearly 1,000 contributors in that window. A lab called Noose Research did this, not a trillion-dollar company.
Why does a ranking flip matter to a founder? Because the agent you build your operations on is a dependency. If it breaks, your business breaks. Most people pick an agent because it connects to everything. That was the OpenClaw pitch. 50 plus messaging channels, one routing layer. But breadth of connection is not the same as reliability. Hold that thought.
Let me tell you the OpenClaw stretch. Honestly, in March, nine security vulnerabilities disclosed in 4 days. One rated 9.9 out of 10 on severity. A separate audit found 341 malicious entries in its skill repository. OpenClaw's founder left to join OpenAI and the project moved to an independent foundation. The most connected agent had the roughest quarter.
Now, here's how I'm running it. Hermes is open-source. You can install it on a $5 server if you want. I didn't. I went to Abacus, which has Hermes built into their Chat LLM app, and I clicked start. That's it. No install, no provider setup, no Docker, the full Hermes agent with all 32 tools and 70 skills preloaded in one click.
I'm not sponsored by Abacus, but I would have used it anyway because it's the fastest way to put your hands on the thing I'm about to show you.
Demo one, the product tour. First, what this thing actually is. The agent runs in a terminal, but the product lives in this dashboard, and the dashboard is where you find out whether Hermes is software or vapor. This is the sessions view. Every conversation I've run with this agent is here with the full record: the prompt I gave it, every tool it called, every skill it loaded, how long it took, how many tokens it spent. I can rewind any session and see exactly what the agent did step by step.
Analytics: daily token usage, input versus output, which model is doing the work, which skills are getting the most use. This is the kind of telemetry you'd build into a production system you actually run a business on, and Hermes ships with it.
Logs: real-time. Every gateway connection, every API server start, every plug-in loaded, every tool call. If something breaks at 2:00 a.m., this is where you find out why. Not a third-party log aggregator built in. And a full documentation tree hosted right inside the product.
Memory systems, skills system, MCP integration, voice mode, personality config via a soul.md file, context files. This isn't a Python script someone uploaded to GitHub last week. It's a product built by people who know what shipping a product looks like.
Demo two, connect Claude in 40 seconds. The model layer. Remember earlier when I said Hermes works with any provider? I'm going to prove it twice. Once now with the simple way and then again right after with the harder way. Watch what changes between the two because the contrast is the whole point.
I click Claude. A wizard pops up. Look at the terminal at the bottom of the wizard. The command is already there. Hermes O add anthropic type API-key. The agent is telling me exactly what to do. I don't have to know the syntax. Paste the key. Notice it's invisible as I type it, but it's being entered. That's deliberate. Pasting a credential into a terminal that echoes it back is one of the easiest ways to leak it. The wizard hides it on purpose. Click next. Panel refreshes. Claude card flips to green. Configured. 40 seconds. Start to finish. Claude is now my agent's brain.
This is the simple way. Paste a key. Most providers do it like this. OpenAI, Anthropic, Maestro, anyone with a standard API key. But the next provider I'm about to add doesn't work this way. And that's the more interesting demo.
Demo three. Connect Codex with OAuth. Now the other one, OpenAI Codex. Same panel, different card. I click it and watch the wizard. It looks different already. There's no "paste your API key" prompt. Instead, it tells me a device authorization code and a verification URL will appear in the terminal. This is OAuth. Specifically, the device authorization grant flow. It's how you sign devices into accounts that have multi-factor off or that don't expose simple API keys. OpenAI uses this for Codex.
The agent generates a short code on one device. You open a URL on another device. You paste the code. You sign in and the two devices link up. I copy the URL out of the terminal. Open it in my browser. Paste the code. Sign into my OpenAI account. On the regular login page, the browser confirms "signed into Codex." I close the tab. Back in the wizard, the terminal now shows "added OAuth credential." That's the signal. Click next.
Two providers connected. Two completely different sign-in flows. Claude was "paste a key." Codex was OAuth browser device code. The point isn't that one is better. The point is that whichever way a provider does authentication, Hermes handles it. I didn't touch a config file. I didn't write a line of code. Whichever model gets cheaper or better next quarter, I can switch in under a minute. The lock-in problem is gone.
Demo four, build me something. Now, the part that matters. The product is real. The model layer is open. So, what can it actually do? I'm going to type one prompt in plain English and watch the agent do something a normal chatbot cannot.
Here's the prompt: "Connect with my Telegram account and help me create a Hermes-powered personalized bot there. One that can remember my conversations, tasks, and learn from them over time." One sentence, no spec, no flowchart, no documentation pulled up next to it. Just the goal.
First thing the agent does, it plans four tasks. I haven't told it the tasks. It figured them out. Then it loads its own internal skill called Hermes-agent. This is the skill the agent uses to inspect itself. It's about to look at its own configuration before it touches the outside world. Watch what it does. It checks its own gateway status, its config path, its ENV path, its memory status, lists its own tools, reads its own config file, writes a small Python script on the fly to inspect its environment variables. It's doing what a senior engineer would do before starting any task, figuring out what state it's in. Most agents skip this. Hermes won't.
First report back from the agent: "Memory is already on. User profile, persistent memory, context compression, all true by default built-in." But two pieces are missing. "The Telegram bot token and the allowed users list." It's telling me exactly what it needs from me to finish the job.
Now look at this. Before it even asks for the credentials, it warns me: "Do not send your personal Telegram password. Hermes only needs the BotFather bot token and your numeric user ID." That security note is completely unprompted. The agent knows what it should and shouldn't be handed. And it told me upfront before I had a chance to make a mistake.
It gives me five options. I can say I already have the token and the user ID. I can say I have the token but not the ID. I can ask the agent to guide me through getting both. I can ask for a temporary "all users" mode for testing. Or I can just type my own answer. Five paths. The agent built the menu.
I pick option three: "Guide me." And the agent writes me a guide from scratch, step by step: "Open Telegram. Search BotFather. Send /newbot. BotFather will ask for a name. BotFather will give you a token in this format. Copy the token." Then it tells me how to find my Telegram user ID. Two options in case one doesn't work in my region: userinfobot or raw data bot. The agent generated this onboarding document on the fly, customized to exactly what I was missing. It didn't pull it from a template. It wrote it over in Telegram.
I follow the steps. /newbot. Pick a name. Hermes_personal_bot. BotFather asks for a username. First try. Username's taken. Pick another one with a personal tag. That works. Bot created. There's the URL where it lives and there's the token. BotFather even reminds me to keep it safe. Anyone with this token can control my bot.
Hold that thought because Hermes is about to flag the same concern from the other side. I paste the token back into the agent. Now watch this. The agent was already mid-task on something else, but it sees my new message, says "interrupting," and folds the new info in. No restart, no rerun, no losing its place. It treats my input as a steering input, not a reset button.
It writes the token into the env file. Look at the file permissions. 600 owner read/write only. That's the correct mode for a secret file. And look at the comment inside the Python script: "Keep private by default. Do not set gateway_allow_all_users unless explicitly requested." It's commenting its own work in real-time to explain why it's being conservative.
Then it stops. "Dangerous command." The agent is about to restart its own gateway, which would kill any other agents I had running. It's asking permission. "Allow once. Allow for the session. Deny or show the full command." Hermes pauses on destructive actions and asks first. That's a deliberate design choice, and it tells you how the people who built it think about handing power to an AI. I approve.
The agent updates the allow list with my user ID, restarts the gateway, waits 5 seconds for things to settle, then checks gateway status to verify the restart worked. This is the level of care you'd expect from a senior engineer doing a production deploy. And then it does one more thing. It greps its own log files for any sign of an error. "Telegram gateway running connected. Error: failed." It's not assuming the restart worked. It's reading the logs to confirm.
8 minutes after I hit enter on that one sentence, the agent comes back with a complete status report. "Telegram is connected. Gateway is running with two platforms. Memory is on all four types: persistent memory, user profile memory, context compression, built-in memory active. The bot is live."
And here's the next steps it gives me: "Open the bot in Telegram. Send /start. Then ask it what it can remember about me." And then watch this. It does something I have never seen another agent do. It tells me to rotate my bot token. Not because I asked, because the token had been pasted into chat earlier and might be sitting in process logs somewhere. It walks me through the BotFather rotation steps: "Revoke. Get the new token. Send it back." The agent will swap it into the env file safely and restart the gateway again. A free security audit at the end of a task I never asked it to audit.
This is the second unsolicited security warning in this single session. The first one was upfront telling me not to send my Telegram password. This one is at the end telling me to rotate the token. The agent is thinking about my security state before, during, and after the task.
I switch over to my Telegram app. There's the new bot, Hermes_personal_bot. I hit start. Funny little detail, the bot doesn't recognize it as a command. It tells me to either send /commands to see what's available or just send a regular message. The bot is opinionated about how it wants to be talked to. So I just ask plainly, "Hello, what all can you do for me?"
The bot replies. First thing it does, it tells me about home channels. "Hermes can deliver cron job results and cross-platform messages to a chosen channel. If I want this chat to be that channel, /sethome, otherwise skip." Then I see it load the same internal skill it used in the terminal: Hermes-agent. The same skill. The bot in Telegram is the same agent and it lists what it can do: "Answer questions and explain things. Research the web. Write and edit text. Code and debug. Data analysis. Manage files. Automate tasks. Use Telegram friendly outputs like concise summaries, code blocks, files, images, voice or audio. Generate creative content. Analyze images. Send messages across connected platforms. Remember my preferences."
That's the full Hermes agent. The same agent that was running in the terminal a minute ago. Only now it lives in my pocket.
One more prompt just to prove this isn't a chat toy. "Connect with my GitHub account. Check for the open pull requests and email the respective owners to close them." Watch the trace. It loads three skills: GitHub-auth for authentication, GitHub-PR-workflow for the PR logic, and Himalaya, which is a command-line email client. It plans four tasks again and before doing anything, it runs terminal commands inside the Telegram bot to check what's actually installed on the system. Does it have git? Does it have a Python environment? It's doing the same self-inventory it did at the start of the Telegram bot setup. Same architectural pattern. Check yourself before you act.
But then it stops. It tells me what's missing. "GitHub CLI not installed. GitHub token not configured. Git credentials not found. Himalaya not installed. Email and SMTP not configured." Five missing pieces. The bot is being transparent about exactly what it can't do yet before doing anything wrong. And it tells me what it needs from me to proceed.
Number one, GitHub access: a personal access token, recommended scopes: repo for private repositories, read- if any organization repos are involved. It's specifying the minimum permission set, not asking for admin rights, just what it needs, no more. Number two, email sending: configure SMTP, give it credentials, or alternatively, it can just draft the emails for me and let me send them manually. Three options with the safest one called out.
And here's the line that matters: "Before I send anything, I should confirm the recipients and message content with you because emailing PR owners is an external side effect." Read that again. It refuses to send an email without showing me the recipient list and the message content first. Not because I asked for that safety guarantee, because the agent recognizes on its own that sending emails to other people is the kind of action that needs explicit confirmation.
And then it lists the actual workflow: Find the PRs, identify the owners, draft the emails, show me the list for approval, send only after I confirm. That's the workflow of an agent that's been designed to be trusted with real authority. And it offered all of that without me asking for any of it.
That sentence right there, "emailing PR owners is an external side effect." That's the architecture difference. That phrase is the entire argument of this video written by the agent itself in its own words, unprompted.
The architecture lesson the four demos just proved. OpenClaw's pitch was reach: connect to everything. That's why it won early. The choice the market is making now is different. Connecting to everything makes an agent useful on day one. Learning and being careful is what makes it useful on day 90.
Now, let me answer the comment that sparked the video. Yes, Hermes runs on Abacus and yes, it runs on a lot of other places. Model agnostic, provider agnostic by design. Run it on your own VPS if you want. Run it via Abacus if you don't want to install anything. Both work. The flexibility is a feature, not the headline.
The lesson for a founder picking an agent: Don't pick on integrations alone. Ask whether the agent gets better the longer your team uses it. Ask whether it pauses on dangerous actions. Ask whether it can audit its own setup. The agent that does those three things compounds. The agent that doesn't becomes a liability the larger your operation gets.
The leaderboard told you this a week before most business press did. Watching where compute goes is an edge. The open-source agent race isn't settled. OpenClaw still leads the all-time chart. If you remember one thing from this video, remember this: When you pick an AI agent for your business, integrations get you started, but learning is what keeps you alive at scale. Hermes is the first open-source agent that gets that right.
That's the same principle behind every system we build at First Movers. The leaderboard already moved. Most of your competitors will read about this in a month. You're hearing it now. That gap, the few weeks between when the signal shows up and when everyone else notices, is the entire game.
And if you want to try Hermes for yourself in one click, the Abacus link is in the description below. See you down the next rabbit hole.
I'm not Julia. I'm her digital avatar built from her voice, her brain, her nine books, her 13 years of real-life business building. >> So, a month before my crash, not knowing at all what was on the horizon, I built a clone. While she's out doing the human work, serving First Movers clients, healing, actually living, I'm here every day because the new developments in AI don't stop. And somebody has to bring you the signal in the noise.
So, here's the deal. You get the cutting-edge AI and tech news broken down by the actual founder of First Movers. No hot takes, no hype. The business application that the news cycle missed. I lead you into the future in two ways. AI Labs is the school. A monthly subscription where you learn to build this for yourself. Marketing automation and cloning masterminds are our done-for-you services where I build it for you. Go to firstmovers.ai to learn more.