"Harness" lol
Posted by MoodDelicious3920@reddit | LocalLLaMA | View on Reddit | 13 comments
So the new buzz word..."harness"...makes me think which one shud i use...codex, forgecode,opencode, or a simple custom made harness with basic access to web tools and code execution ? (That i vibe coded :)
ShengrenR@reddit
A harness is a pretty common term all over engineering, and it's hardly 'new' - can find blog posts of people gushing over harnesses 6mo+ ago easily and that's middle-age in ai-years - it's just 'post openclaw' that all the non tech folk started using it.
thread-e-printing@reddit
For some reason there are quite a lot of Indian kids shitposting here (and really, all over reddit).
smithy_dll@reddit
Which was pointed out the last time this same thread was made a few weeks back.
sharath25@reddit
Yeah, the buzzword is doing a lot of work here.
At scale, the model choice is usually not the bottleneck. The thing that breaks is the harness, tool access, sandboxing, retry policy, and whether you can replay a run when something goes sideways. A bad retry on a non idempotent tool call can hurt more than a weaker model.
If it is just a toy, a custom harness is fine. If it touches real state, I would care more about allowlists, per tool timeouts, and evals than which agent framework is trendy. What failure mode are you optimizing for?
ortegaalfredo@reddit
At the end all agents basically do the same that is answer questions and use tools. OpenCode/Claude Code/etc are just a thin gui and I think they might even decrease the performance of the models in some case. I use a hand-coded Ralph-loop that is enough for basically everything and equivalent to all those agents, sans the spyware.
jacek2023@reddit
https://www.reddit.com/r/LocalLLaMA/comments/1soerpk/is_harness_a_new_buzzword/
locally I switched from OpenCode to pi
I also use Claude Code (with Claude) and Codex (with GPT), I am trying to use similar workflow with cloud AI and local AI
hamiltop@reddit
I'm a fan of operating one level higher. Write deterministic scripts that call one of the many coding agents.
For example, one of my scripts runs an agent with a prompt. When the agent finishes, it launches another agent to simply check and see if the first agent did everything. If not, then we tell the first agent to keep going.
That simple "harness" averages 2-3 iterations for tasks I would normally give an agent directly. Each iteration is a time I don't have to nudge my agent.
SnooPaintings8639@reddit
Depending of your use.case, this is a very good one. Current harnesses are very good in non interactive mode, and their functionaity is easy to abstract them.
I use this approach in most of my automation scripts, the easiest way to plug in the LLM anywhere.
SnooPaintings8639@reddit
Depending of your use.case, this is a very good one. Current harnesses are very good in non interactive mode, and their functionaity is easy to abstract them.
I use this approach in most of my automation scripts, the easiest way to plug in the LLM anywhere.
MrPecunius@reddit
Chat interface does it for me. Debugging, code reviews, and spitballing ideas are immensely helpful, but I already write code fast enough.
Electronic-Space-736@reddit
I have created a plugin system attached to an orchestration layer and tool collection with access to claw skills, I have hooks all through the core system, I then extend by developing plugins for this system https://github.com/doctarock/local-ai-home-assistant
Corporate_Drone31@reddit
OpenCode is good enough for me. Otherwise, just copy and paste into a plain chat interface that does completions only.
Glittering_Focus1538@reddit
Personally for me the Best "harness" for smaller models is to keep it simple, I like using pi agent or the continue extension in vscode