favorite Agentic Coding Harness

Posted by chibop1@reddit | LocalLLaMA | View on Reddit | 28 comments

So far, I’ve tried Codex CLI, Claude Code, Gemini CLI, OpenCode, and recently, Pi with local models.

Pi is the leanest of them all, with just four tools: read, write, edit, and bash. Its system prompt is only under 2K tokens, and it's perfect for local models.

I've been trying out Qwen 27B-MXFP8 with it, and it's much better than I expected!

It doesn't have fancy bells and whistles like multi agents, but the only thing I’m missing is searching the web for documentation. I’m sure you can get it through an extension, but you probably won’t get the same robust search features you get from commercial platforms anyways.

What’s your favorite agentic coding harness?

[-]

Creative-Type9411@reddit

im making everything custom i turned off built in openwebui search and just have a custom HTTP request tool for direct interaction, API, etc., and something separate for web search that browses things depending on what I'm looking for

i just finished a reddit search tool

amazon/newegg/ebay(waiting for dev api)

Instead of using someone else's tool, I'm making powershell scripts to send requests and hand parsing the returns to be able to make them reliable.. then once I get the formatting down properly and any workarounds I need I converted over to python openwebui tool format

[-]

DeltaSqueezer@reddit

I think this is the future. Everyone will create their own custom UI that suits how they work.

[-]

alexmrv@reddit

100% I plugged in my psychometric report to Google stitch and asked it to design a UI for my particular way of processing, I got a UI that has the _perfect_ information density and context-switching workflow for me in ways I didn’t imagine. Have it export the UI prototype and hand it over to pi. One-shot prompted my own tool on 21 usd of tokens

[-]

alexmrv@reddit

Pi all the way.

I think llm workflows are fundamentally ungeneralizable (?) I’m not sure this is the right term but what I’m trying to say is: the same setup for two people will not necessarily work for both of them because the way you think influences the way you talk and that changes how you write and how you write means what tokens go into the pipe, and the outputs depend on that.

Also projects are different meaning the same tools will yield different results at different qualities, we’ve been struggling to just have a benchmark of what works and some of the best researchers in the world are trying a failing to answer the question: is this good?

All this to say: anything but build-your-own is just hoping you can adapt to the style of the people that made it.

Plus, making your own starting from Pi is like… a couple of afternoons of work.

[-]

LetsGoBrandon4256@reddit

How much a scrub am I if I use the copilot extension in VS Code?

[-]

Risen_from_ash@reddit

I use Cline in VS Code and it’s absolutely incredible.

[-]

yes2matt@reddit

... can't get no love ...

[-]

Limp_Classroom_2645@reddit

Pi coding agent, i love simplicity

[-]

lioffproxy1233@reddit

Host searxng locally and have pi build a skill to use that port.

[-]

JamesEvoAI@reddit

I've been finding great success with little-coder, which is just Pi with some extensions to help shore up the common issues of local models:

https://itayinbarr.substack.com/p/honey-i-shrunk-the-coding-agent

https://github.com/itayinbarr/little-coder

[-]

BidWestern1056@reddit

npcsh is similarly lean like pi

https://github.com/npc-worldwide/npcsh

[-]

son_et_lumiere@reddit

aider-desk

[-]

DiscipleofDeceit666@reddit

Qwen code was the first one that worked with the lobotomized version of qwen3.6 35B moe

[-]

fractalcrust@reddit

oh-my-pi

[-]

CardinalRedwood@reddit

How’s it different than pi? Seems sick!

[-]

HomsarWasRight@reddit

I think it’s basically just a pre-configured Pi agent. Pi comes with no bells and whistles built in and expects you to choose extensions and configure it yourself.

Personally that’s why I like it, but this version does seem nice.

[-]

CardinalRedwood@reddit

Got it. Ty!

[-]

eikenberry@reddit

Goose has been adopted by the AAIF as a standard agent. I've used it a decent bit and it is pretty good. https://goose-docs.ai/ https://github.com/aaif-goose/goose

Another promising looking agent is Crush.. haven't played with it as much but it is on my list of ones to try. https://github.com/charmbracelet/crush

Another new, lightweight one I came across recently is zerostack. Not as well known as the others (github star wise) but looks promising. https://github.com/gi-dellav/zerostack

[-]

NoWorking8412@reddit

I love Pi. My first coding harness was cursor, shortly followed by Claude Code, where I've been ever since. I've tried every other harness that I've come across, especially in an effort to move away from Anthropic and go more local, and Pi is the first one I've felt that can replace Claude Code for me.

[-]

Juulk9087@reddit

I made my own.

[-]

Global_Tap_1812@reddit

I'm accidentally doing this lol. It started off as a model handoff where one model prepares context and creates an optimized prompt (and sets proper security permissions for subagents) and now I've just realized I want it to be a unified CLI where I can jump between models like opus, codex, Gemini, and Qwen, and have some more structure to the workflow

[-]

Pleasant-Shallot-707@reddit

Pi coding agent

[-]

Expensive-Shift8584@reddit

I'm using hermes currently. As a novice to all of this, I like that I don't have to set up much up, it can figure most things out on its own and writes a skill if its a repeatable task. I started a few months ago with Aider, then tried out claude cli powered locally, and eventually antigravity with the llm powered as an mcp. Hermes has been my favorite to work with, developing simple html5 games with my kids over telegram.

Still trying to figure out which model is best. I've been using Qwen 3.5 397b at q4 on my mac studio m3 ultra 256gb, seems pretty fast and accurate. Also have minimax 2.7 q6 in the roster, as well as all of the gemma 4 and qwen 3.6 models, but they get less use

[-]