News for hackers. Stuff that matters.
Re:Ask HN: What's your current AI agentic setup? by cdnsteve on Wednesday November 04, 2026 @10:52PM (#50374953)
I've done the long loop in the past year and tried a large variety of models, agent harnesses and desktop versions. I actually don't like any IDE/agents mixed together (VS Code). I prefer a new experience.

First Claude Code CLI, was power user, loved it a lot. Was paying for their uber max 20x big daddy plan for a long time. Built tons of shit, github was all green. Shipped like 10 web apps, 1-2 smaller SaaS products, an AI community site, tokenstead.ai

Tested codex but wasn't really a fan, ever.

Heard about OpenCode, tried it out. Wasn't enjoying the harness as much. Tried PI briefly but all the plugin confusion and user addons made me nervous for getting malware. Tried Copilot CLI and wasn't impressed, felt super early and immature, buggy. Briefly tried OpenClaw, it seemed like a security nightmare so uninstalled in like 3 days.

Started testing a bunch of model providers/inference and open-weight models, Ollama, Openrouter and model/harness ones like OpenCode.

I used Claude Code Harness with Ollama running models as my daily driver for awhile. I started using Hermes with Ollama models. Hermes Desktop, and Hermes running on a PI5. Built a bridge between them, setup a voice model on a Jetson nano origin for Hermes. Connected to Telegram.

This was super cool and useful, hooked into Sugar, my own open-source memory tool for all these AI tools, 100+ Github stars.

I tried our Herdr and T3 Codes. These are also very interesting variations of Desktop/CLI experiences. Worth testing if you haven't given them a shot.

What I recently saw at Github Universe makes me think Copilot Desktop could actually take over most Enterprise experiences for harness. Testing this on my personal machine. It seems to have borrowed a lot from T3 Codes and other Desktop harness tools. Most companies are using Github and Copilot already so this is dead simple to bring in. The last mile are things like sandboxing, governance, etc.

Right now I'm giving Deepseek Harness a spin. For the past few months I've been trying to move away from frontier providers, models and products. Open-weight should win and so should local AI.

In summary there is a massive amount of churn in this space. Trying to plan on the best tool is very difficult and frustrating, the SDLC is a moving target right now. What we need are better systems for development.

Building actual products and then trying to MCP/agent everything for sales, is quite painful. Apollo this and whatever tool that, you just end up spamming everything crappy cold emails.

I'd be very careful to fall in the trap of trying to build this system yourself. We're all builders and it's easy to have visions for a perfect system and process. Keep an eye on what the hyperscalers are doing for running agents in prod, a different but related space. I have a feeling these will all converge and everything will just be cloud based, real-time agent factory thing.

With all this said and done, I think we're actually regressing. The web use to be a social place, now it feels like everything is AI slop and finding genuine connections is lost. ICQ says, Messenger days, IRC, message boards. I met some amazing people, exchanged ideas and friends I still know. But now, the whole social web is toast. The AI spam that people post is everywhere, no one reads anything anymore and we're all overloaded with things so we use our own AI to try to make sense of it.

I can honestly say after doing this for about a year full time is I am nearing AI burnout. No human should context switch every 4 minutes between 7 active agents, mostly running different products.

I might just become farmer in a few years. That or open up a tiki bar on the beach somewhere and call it early retirement.

/fin

There are no comments on this comment yet.