Agentic AI Isn’t Just Coding

Agents didn’t click for me until one updated my website.

Claude Code already felt like a jump from chatbots: give it a task, let it edit the files, review the diff. That’s agentic coding, and it’s great. But it’s still you sitting at the keyboard, driving.

What changed my mind was watching Hermes Agent handle everything around the code.

I’d been skeptical of OpenClaw since it launched. Lots of hype, no receipts. People posted about using it but never showed what it was actually good for, so I never bothered deploying an instance.

Now I keep hearing about Hermes, a refined agent similar to OpenClaw with self-improvement and persistent memory, so I figured I’d finally try it for myself. The whole point of this blog is to see whether AI tools are worth the hype, so I deployed one.

Here are two workflows I’ve actually been using.

A daily AI news roundup

Instead of me manually checking Hacker News, Twitter, newsletters, GitHub, and product launches every day, Hermes does the first pass.

The trick is the filter. I don’t want “AI news.” I want builder-relevant updates: new developer tools, coding agents, local models, open source projects, browser APIs, small workflows, weird tools worth testing. I don’t want funding announcements, vague enterprise partnerships, benchmark hype, or big company strategy posts unless there’s something I can actually use.

Hermes runs on a schedule, filters the noise, and only pings me if something clears the bar. If nothing’s worth reading, it stays quiet.

That last part matters. A normal feed reader still makes me decide what matters. Hermes acts more like a research assistant: it knows what I care about, checks repeatedly, and only interrupts when something is worth my attention.

Updating my Astro site

My blog is an Astro site. Normally, if I want to change something, I launch a session on Claude Code and have it update the files, run the build, commit, push, open a PR, wait for Netlify, and then check the preview.

Not hard. Just a lot of little steps that are especially annoying if I’m doing them from my phone.

With Hermes I can say “update this page on my site,” and it also makes the change, but it also runs the build, pushes the branch, opens a GitHub PR, waits for Netlify, and sends me the preview link.

That was the moment agents clicked for me. Not that it could write code — Claude Code already does that. It was that Hermes could handle all the work around the code. It felt less like chatting with a bot and more like delegation.

Orchestrators, not chatbots

When I ask Hermes to update something on my site, it isn’t doing the coding itself. In the background it’s launching OpenCode and prompting it to make the changes. If the build fails, it goes back to OpenCode until things resolve.

Because the agent has a little autonomy, I don’t have to walk it through every step. I don’t ask it to search for news, then filter it, then summarize it. It does all of that from a single prompt — and it remembers how.

This is also why I think a lot of the agent conversation feels incomplete.

It usually goes one of two ways. First, demos: an agent opens a browser, clicks around, books a fake flight, orders a pizza. Flashy enough for a video, but most of us aren’t booking flights often enough for that to matter. Second, security: what if the agent deletes files, leaks secrets, spends money, gets tricked by a malicious page?

The security conversation is necessary. If an AI can act, then permissions, approvals, logs, and sandboxing matter a lot. But if the conversation stops there, we miss the more interesting question: what are these things actually useful for?

My answer, right now, is that agents are useful for the workflows that sit between intention and completion. The annoying chain of steps between “I want this changed” and “here’s the thing ready for you to review.”

That might be updating a website. Monitoring a topic. Cleaning up a repo, drafting a PR, checking a deploy preview, reminding you when something changed. The value is closing the gap between asking and doing.

We’re still early. A lot of this is fragile. You have to be careful what you give an agent access to, put approvals in the right places, read the diffs, check the previews, and treat the agent like something that can make confident mistakes.

But I no longer think of agents as chatbots that can code. An agent can carry a task across tools, time, and context.

That’s the part I didn’t understand until I saw it happen on my own site.

The future of agents won’t be defined by how impressive the demos look. It’ll be defined by whether people can trust them with the boring workflows they already do every week.