My Emacs Agentic Workflow

Two years ago I was a skeptic about GenAI. Now Emacs has become my agent cockpit.

Harness

My favorite harness right now is hermes-agent, which is probably the “emacs of agent harnesses” with how easy it is to configure and extend, while not being “just an agent harness.” Hermes supports messaging gateways, so you have first-class support for mobile as well. All you have to do is /handoff <platform> to continue your session via a mobile messaging app.

To interact with hermes via emacs, I’m using emacs-hermes, which is available via NonGNU ELPA.

Sometimes I also use codex via emacs-codex-ide. But it’s not as token-efficient as my hermes setup, and it’s not where I am most productive. Unlike most harnesses, hermes supports codex native compaction, which is the “killer” feature of codex.

Models

I currently use gpt-5.6-sol for most tasks through my codex subscription. It is exceptionally good at long-running goals. It can easily keep going for 20h+ trying to complete its task, and I have found the quality of its work more than adequate.

TIP: to save tokens with hermes, adjust the auxiliary.approval model to a cheap open-weight one, such as deepseek-v4-flash, which is practically free on nous portal. This has cut my usage by more than 35%.

I’m also surprised by the recent leap x.ai made with Grok, which is currently the best value-for-money model by far. The benchmarks place it close to gpt-5.6-sol and fable.

It’s currently my fallback model when I reach my usage limit on codex. It’s the fastest “frontier” model I’ve tried.

I avoid using models from Anthropic. Models like opus-5 and fable are good for coding tasks, but I found them subpar for my studies even compared to grok-4.6, plus overpriced. The API prices are insane, and their subscription only supports their proprietary harness.

What about open models?

A friend and I tried to self-host our “AI Homelab”, repurposing an old ethereum mining rig plus a couple of new Nvidia GPUs. The performance was not ideal with models like GLM-5.2 compared to subscriptions by a long shot. Forget subagent workflows if you ever attempt it. Not to mention that my electricity bill was the same as I’d pay for a subscription plan.

A lot of people in the “GNU/Linux” sphere are looking for “free” models. Even if there were a 6T+ “free” model that outperforms frontiers, there is no “free hardware” that could run it, and even if there were a FOSS Nvidia alternative, would it be financially worth it for most people as an investment? Would your average hobbyist rather pay $20/month or buy the next gen “free” DGX Spark for $5,000?

Extra goodies & Skills

Memory

No matter your setup, the biggest upgrade for me has been hindsight. It’s an agent memory system that you can share across harnesses, and it helps your agents remember your preferences and workflow details.

It’s free and easy to self-host.

Skills

Skills are where my setup stops being generic. Instead of explaining the same preferences and workflow every session, I put them in SKILL.md files that Hermes loads depending on the task. I have skills for how I want elisp written and reviewed, and even for drafts like this post.

Everyone takes a different approach, from GSD/BMAD/Spec-driven workflows that automate most of the work to “real engineering” skills that want you to babysit the agent. I prefer smaller skills I can mix together instead of one framework owning my whole workflow.

If you are just starting out, browse skills.sh for anything that “looks cool”, experiment with it, then start writing your own. The two I recommend to most people are unslop and grill-me.

Other tools and environments

As of now, the only other tools I’d recommend are web tools like firecrawl and searxng.

Current state of my “AI Psychosis”

The past year the value I have gained from GenAI tools is insane. I had not dreamed, even in my wildest dreams, that I would have such tools in my lifetime.

Tasks that used to bottleneck my workflow, such as adjusting a software tool to help me study, are now cheap to solve. Software is more of a commodity now, no longer a moat. Anything you like can be built easily. You can see that even in projects like hermes, which copied the “Grok Bot” “killer-feature” in less than 24h and made it open source in their harness.

Grunt work like notes for my syllabus based on my university lecture slides used to take me weeks. Now it’s just a /goal prompt, go for a walk for 2h, and come back and they are ready in the same exact format the rest of my notes already have. All that is left is for me to study them.

We live in exciting times.