Every week there was one or the other demo. An agent booking flights, replying to emails, researching competitors, running entire workflows on its own. The goal was always the same, stop opening ChatGPT every time you need something, and start handing work to an AI that keeps going long after you’ve closed your device.

It sounded great but actually building one was a different story.

The model was never the hard part but everything around it was. APIs, hosting, scheduling, memory, monitoring, deployment. The models became dramatically smarter but the workflows barely changed. Most people still use Claude or ChatGPT exactly as they did a year ago.

A few days ago, Anthropic released something called Launch Your Agent, a GitHub repository that installs as a skill inside Claude Code.

Img 1: The Github Repo

At first glance it doesn’t look like much. After spending a few hours building with it, I feel it’s one of the more consequential things Anthropic has shipped this year. Not because it removes every technical hurdle. Because it changes where the effort goes. Until now most of the effort was spent assembling the infrastructure. This shifts it toward defining the work itself.

For me running a YouTube and Instagram channel about AI means staying current is part of the job. Every morning the same research ritual plays out before a single word gets written. Reddit, GitHub, YouTube, research papers, product launches. One question: what’s worth paying attention to today?

That isn’t creative work. It’s recurring research. And recurring research is exactly the kind of task an agent should handle.

So the task was simple: actually build it, end to end.

Here’s what to do if you want to follow along. Go to link.

Copy the repo link, open Claude Code inside Claude Desktop, paste the link, and say: install this skill. The surprising part wasn’t how quickly it started generating files. It was how little it cared about writing code at first. Instead, it asked questions.

For my case, it asked questions like: Should the agent optimize for breaking news or evergreen ideas? How recent should a source be to count as relevant? How should it know when an idea has already been suggested? What makes a morning report useful enough to actually read rather than ignore after the first week?

Image 2: Claude Code asking permission to clone the repo. The permission prompt when installing. Choose Allow Once the first time, not Always Allow.

Most people assume they know what they want until someone forces them to write it down. That’s what this skill does. It made the intake conversation happen before anything else could proceed.

And the cost of vagueness here is different from a normal chat. In a regular conversation, vague answers produce vague output and you iterate. Here, vague answers produce a vague agent running automatically every morning

Once the workflow was clearly defined, Claude produced a full plan before writing a single line of code. Model, tools, memory, evals, output format, schedule, delivery. Everything laid out before anything was built.

That’s the part most people skip when building agents. They jump straight to code and figure out the rest later. This forced the thinking to happen first.

Image 3: Asking clarifying questions: The scheduling intake conversation showing 7am IST, India time zone, last 7 days research window.

For this agent: Sonnet was used as the model, so the 7am schedule fires whether or not the laptop is open, web research as the primary tool, and output as a dated markdown file every morning.

Add a memory store. Without it, every run starts fresh with no knowledge of what came before. The agent will hand you the same five ideas every Monday. With it, the agent stores every idea it generates and checks that store before producing new ones.

The same logic applies to evals. An eval is a quality contract for every run: what must the report contain, what counts as a pass, what counts as a fail. Answer that before you build. It becomes the standard the agent holds itself to on every run.

Once the plan was confirmed, deployment was straightforward. Claude generated the project structure, hit an error partway through, diagnosed it, fixed it without intervention, and pushed everything to Anthropic’s Managed Agents platform. A year ago that error recovery alone would have meant an hour of debugging.

Note: the Anthropic API key goes into the .env file in the local agent folder directly, not into Claude Code itself. Then tell Claude Code the keys are updated and to proceed.

Before anything goes live, open the agent-overview.html. It shows the full picture: deployment details, environment, tools, memory, evals, output format. Review it. The plan exists at this point but it hasn’t launched yet.

Review it before deploying. The plan exists at this point. It hasn’t launched yet.

Image 4: The agent-overview.html showing the full plan.

After deployment, the work moved to Managed Agents, where agents get observed rather than built. The Observability tab shows every run, every failure, every token consumed. Without it the only way to know whether the agent is working is to trust that it is. Don’t trust. Check.

One lesson from running it: first instinct was Opus. Five dollars per run, thirty-five dollars a week. Switching to Sonnet produced almost identical output quality at seventy percent lower cost. Start with the cheapest model that clears the bar. Upgrade only when output genuinely suffers.

Anthropic hasn’t eliminated the complexity behind production agents. Models still need selecting. Workflows still need designing. Evals still need writing. But enough friction has been removed that the first question is no longer “can I build this?”

It has become “what work do I want to run before I wake up tomorrow?”

That’s a much better question to be asking.

You must have seen it. Someone posts their agent setup on Twitter, gets 3,000 likes, and three weeks later they’re back to just using ChatGPT like everyone else.

It’s not because they built it wrong. It is because the agent didn’t know them. Every session started from zero. Same context, same re-explaining, same generic outputs. After a week of that, the novelty wears off and you go back to what’s familiar.

The agent wasn’t growing with them. It was just running. Don’t get to the numbers directly

Hermes was built to solve exactly that. The core idea is stated right on their page that it’s an agent that grows with you. Every conversation you have, it updates its memory. With every interaction it gets sharper on who you are, how you think, what you prefer, how you want research done.

Before Hermes Desktop was launched, my team was running Hermes locally and on a VPS, using Telegram as the interface. And honestly, it worked. You’d tell it your requirements for an Amazon purchase and it would go research, compare, come back with a real answer.

But it wasn’t something you’d want to use every day. You had to pipe everything through Telegram, it wasn’t efficient, and it was hard to communicate with cleanly. Worse, since nobody creates multiple sessions in Telegram, everything piled into one chat

People want a proper UI where they feel in control. That’s exactly why Hermes Desktop came out.

Download page, Mac, Windows, Linux. Everything runs locally. Your data stays on your machine.

Hermes Desktop is available for Mac, Windows, and Linux. Once you download and install it, everything runs locally - your data isn’t going anywhere. You connect it to a model to make it run. Hermes has their own, but you can also use OpenRouter or several free APIs.

For my setup, I used OpenAI Codex, (the $20 plan). Anthropic’s Claude subscription wasn’t used because Anthropic is strict about using their subscription for these kinds of agents. Agentic usage is subsidized and requires a lot of tokens. OpenAI is more open to it, which is why Codex was the choice here.

Worth noting, Claude does have memory, it keeps improving, and Claude Code and Cowork are genuinely agentic. They can interact with local files too. But Hermes gives more control. You control the conversations, the skills being stored, the search happening from your local environment.

And speaking of skills, this is where Hermes starts to feel genuinely different from anything else I’ve used.

A skill is a reusable prompt that makes the agent behave like a specialist in a specific area (you must have seen it used in Claude!). Hermes ships with 83 by default. I have added custom ones for content writing, and our brand voice. When I ask it to draft a post, it already knows our tone, our format, our audience. I don’t explain every time.

Skills library, categories across the top, individual toggles per skill. Only 2 are ON here. Everything else is off.

One rule though, don’t enable all 83. Every active skill’s title and description gets loaded into every prompt you send. It’s not a massive cost but it clutters the model’s attention. Only switch on what you’ll actually use. If you haven’t touched a skill in three days, turn it off.

But skills alone aren’t enough. Skills make the agent smarter. Toolsets make it actually do things, web search, browser automation, computer use, cron jobs, task planning, image generation, video analysis, file operations.

Toolsets tab, 20/27 enabled. Cron Jobs, Computer Use (macOS), Browser Automation, Code Execution, all live.

Tools make it capable of doing things on your behalf instead of just saying things. That’s the difference between a model and an agent.

And then there’s profile, a completely separate agent configuration, its own model, its own system prompt, its own persona, its own skill set and its own behaviour.

I have one profile running Claude specifically for brainstorming. I go there only when I want to think (different model, different context, different output style entirely). My research profile has web search maxed out and a tighter system prompt. My content profile has all our brand skills loaded.

You can’t do this cleanly with Codex, Claude Code, or most other tools by default. You’re always using one version of the same thing. Profiles give you specialisation without rebuilding context every time you switch tasks.

The model picker, every major provider in one dropdown. Auxiliary tasks (Vision, Web extract, Skills hub) can run on a different model from your main.

One thing to know before you set up Hermes Desktop - it has full local access to your machine. It understands your file structure, works with local documents, automates things that would otherwise need manual effort. But one badly worded instruction can become a problem. Be deliberate about what you ask it to touch. If that makes you uncomfortable, run it from a server via the remote gateway instead. Local means data never leaves your machine but cron jobs only fire when your laptop is open. Remote means always-on but data lives on the server. Pick based on your use case.

The real value here isn’t any single feature. It’s that you can finally see the whole stack, memory, skills, tools, schedules, models, and control all of it. Claude and ChatGPT aren’t bad. But Hermes puts you in charge of how the agent runs, what it knows, and what it does on your behalf.

Most agent setups don’t survive the first week because the agent never learned who the user was. With Hermes, that’s the whole point. It’s not just running tasks. It’s building a picture of you, every single day.

Download Hermes Desktop. Set up one clean session. Enable only the skills you’ll use this week. Run one cron job. See what comes back.

Leave a comment