Dang thats a huge insight, just what i needed right now :D im having 16gb vram and decided to cancel all my subscriptions for a while to try local models since im having 80-90t/s on qwen 3.6 35b. I want to try some of these harnesses (i finally understand what they are haha) i already tried gemini cli, codex, claude code, opencode, now im playing with PI (which is super cool, im implementing steel browser with screenshots to help me "SEE" stuff instead of be blind :D) but im really thinking about aider now...
also i was surprised you didnt mention clawbot or hermes agent, but now that i think of, they are not just harness, they are full stack so they kind a different subject here. anyway amazing article
PS: zawsze wiedzialem, ze polacy maja dryg do bycia w czolowce AI :D
Hi, thanks for the comment. The Hermes and Openclaw are actually the frameworks of the AI agent, so they contain architecture scripts and everything like that. Packages and Harness is kind of different, but I can see that you understand that.
Also: Tak, Polacy mają dryg do technologii! Generalnie Polska jest mega nowoczesna pod kątem cyfryzacji.
1. Harness is what connect LLM to terminal, code, tools.
2. LLM only outputs text. Harness use LLM outputs, execute, get results, append to text to give input to LLM
3. Harness is as important as LLM itself, without harness LLM is very weak in real tasks. You can use 1 LLM with 2 different harnesses to see results
4. 2 types of harness: code helper, and Agent orchestrator
5. 2 most popular harnesses: Claude Code & Codex CLI
6. There are many third party harnesses
7. Harness is so important, that’s why companies lock users to only use their LLM with their harness (like only use Claude with Claude Code, otherwise pay more)
8. Some research shows that meta-harness (kind-of genetic learning to self-improve harness) is promising
9. If there is problem, look at harness before looking at LLM itself.
10. Context lenght is the only real constraint. Others can be tuning in harness.
ok, I'll take a look at PI. I could not tell from your post, initially, what your full impression of PI is. I'll go give it visit.
BTW - side note, despite my best efforts at reducing token burn in my writers assistant, I can see i'm bumping into limits and wanting to use the Max plan. I haven't put something in place as you have done so well of creating income with mine but I'm giving that some thought. Anwway, thanks for charging ahead and learning about these other tools. So helpful.
Dang thats a huge insight, just what i needed right now :D im having 16gb vram and decided to cancel all my subscriptions for a while to try local models since im having 80-90t/s on qwen 3.6 35b. I want to try some of these harnesses (i finally understand what they are haha) i already tried gemini cli, codex, claude code, opencode, now im playing with PI (which is super cool, im implementing steel browser with screenshots to help me "SEE" stuff instead of be blind :D) but im really thinking about aider now...
also i was surprised you didnt mention clawbot or hermes agent, but now that i think of, they are not just harness, they are full stack so they kind a different subject here. anyway amazing article
PS: zawsze wiedzialem, ze polacy maja dryg do bycia w czolowce AI :D
Hi, thanks for the comment. The Hermes and Openclaw are actually the frameworks of the AI agent, so they contain architecture scripts and everything like that. Packages and Harness is kind of different, but I can see that you understand that.
Also: Tak, Polacy mają dryg do technologii! Generalnie Polska jest mega nowoczesna pod kątem cyfryzacji.
10 things I learn from this post:
1. Harness is what connect LLM to terminal, code, tools.
2. LLM only outputs text. Harness use LLM outputs, execute, get results, append to text to give input to LLM
3. Harness is as important as LLM itself, without harness LLM is very weak in real tasks. You can use 1 LLM with 2 different harnesses to see results
4. 2 types of harness: code helper, and Agent orchestrator
5. 2 most popular harnesses: Claude Code & Codex CLI
6. There are many third party harnesses
7. Harness is so important, that’s why companies lock users to only use their LLM with their harness (like only use Claude with Claude Code, otherwise pay more)
8. Some research shows that meta-harness (kind-of genetic learning to self-improve harness) is promising
9. If there is problem, look at harness before looking at LLM itself.
10. Context lenght is the only real constraint. Others can be tuning in harness.
Oh dude, you are so obsessed with harness that from your words, harness become Prometheus that brings us the fire (uh oh, the LLM) :D.
It's like 2 months ago you are so obsessed with AI agent memory system that you talk about it everywhere :)
Pawel this is super insightful. Thank you so much for researching this and passing along your insights.
I am going go carefully review go see where i can benefit from for my Writers Assistant.
Thanks Tom! And just to be clear - these are only some of options out there! But…try Pi. It is SO GOOD.
ok, I'll take a look at PI. I could not tell from your post, initially, what your full impression of PI is. I'll go give it visit.
BTW - side note, despite my best efforts at reducing token burn in my writers assistant, I can see i'm bumping into limits and wanting to use the Max plan. I haven't put something in place as you have done so well of creating income with mine but I'm giving that some thought. Anwway, thanks for charging ahead and learning about these other tools. So helpful.