I know this is getting meta, but I am sure people and AI would learn a lot from this log:

https://gist.github.com/ontouchstart/b6a08b068cf863f0523bbd8...

The amount of times I hear people running without a sandbox is just crazy!

I'm just looking into setting up a separate GitHub user to let claude act as a contributor, and so much of the docs are about how to prevent your github auth info from your main account accidentally being used. That implies no sandboxing!!

1) none

2) Local LLMs on an M1 Max (Unsloth, LM Studio, occasionally direct llama.cpp). Muse Glimmer (because it is good at what it is optimised for: interactive coding sessions). Qwen 3.8 27B occasionally and a couple of other things for fun (an LFM MoE, experimenting with Nemotron Bolt 74B-9B, enjoying the really super Gemma 4 12B QAT when working on creative photography research or non-code planning).

No openrouter (yet) but that will definitely happen.

3) Currently focusing my attentions on Pi via Paseo (and trying Pendant). Pi is just much better in smaller context sizes, and it is the right fit if you are keen to learn from first principles.

I tried OpenCode first. I couldn’t quite get on with it and what ultimately made me switch was trying to understand its terribly messy source code.

4, 5) I am mostly using AI as a coding sidekick so I don’t really have much need for orchestration yet. I did some experiments with Osaurus (really quite impressive) and I actually think that is almost enough by itself, except that I want a little more control over the sandbox container.

I feel that my usage pattern is still settling, because I am not keen to surrender all of my coding to the machine. But I am trying more purely “directed” coding to learn the process, and finally beginning to get a feel for how I will work with that.

- Antigravity (AGY - Google Ultra $100/month), mostly gemeni flash but I benchmark Gemini Astra sometimes. My main driver because it’s pretty good at search compared to setting something up with DSH.

- DeepSeekHarness (DSH) Desktop using:

—- with pay as you go DeepSeek tokens, more expensive to use during my night/Chinese daytime, otherwise very cheap

—- oMLX on my MacBook driving mostly a Qwen 3.6 35b A3b MOE (mostly used for benchmarking a less capable model)

I mostly do custom subagent programming and can trigger subagents in AGY and DSH easily enough. DSH supports a Python coordinator orchestrating DSH subagents as conversations via RPC. AGY only supports a sub-agent orchestrating other sub-agents but it works well enough at this point. I plan on releasing my subagent framework soon (called autofab in pypi), which is basically a specification driven code development agent system (using existing agent harnesses rather than implementing my own).

I tried goose, OpenClaw, Claude, codex, etc…and they didn’t really work for what I needed. I might get a $100/month ChatGPT subscription to go with my Google one, primarily to get a different perspective and benchmark openAI models on codex (Claude doesn’t really support what I need it to, otherwise would go for that instead).

I have a custom frontend on my android running claude and antigravity CLIs with a UI built on top. Works surprisingly well. It's now about 40 % of my usage . I use it to apply for jobs automatically etc. and orchestrate stuff using adb

I use

- opencode 2 as the harness it works well for pretty much any specific task

- openrouter to use deepseek flash 4.1 … insanely cheap, works amazingly well and fast

- orca as the orchestrator to manage all the work (self hosted). There’s a mobile app to orchestrate afk too

And yep, vscode to browse repos etc

1) ChatGPT Pro ($200) 2) APIs for Bivia (previously used Jev) 3) Oh My Opencode - sometimes, not as much. 4) Herdr 5) Hermes, Openclaw, Dots, Instinct, Pickle. Hermes and openclaw for most routines. Dots and Pickle to call while driving. Instinct because imessage makes it easy on the go.

The somewhat novel part of my workflow, is after Claude has built something out, I have a second fresh Claude session review the code with a set of 16 custom questions. I've discovered that giving Claude questions to ask during code review improves things quite a bit. The questions mirror the exact things I would have asked myself about a pull request when reviewing it. I found this to be a real game changer.

The code comes out quite a bit better, and I would suspect that it uses less tokens in the long run. Simpler cleaner code I would suspect uses less tokens.

I use AI primarily through work, I don't pay for a personal AI sub.

I do have some OpenRouter credits but I mostly use them sparingly to try open weight models out for small personal projects.

I currently interact primarily with Claude Code, but most of my async workers are on Codex.

The orchestration part is custom. I talk to a "chief of staff" Opus instance to ingest various workstreams (Jira tickets, alerts, personal ideas, etc) and turn them into beads. It launches workers to tackle individual beads using Remuda, an agent launcher I wrote at work in 2025 and open sourced earlier this year: https://github.com/yendo-eng/remuda. I configure profiles in Remuda that launch containerized agents using my preferred harness and model. I update these profiles as the pareto frontier shifts. Currently it's GPT 6 Luna max for most tasks.

It monitors the workers and has them go through at least one round of agentic code review. It surfaces PRs for me to read once they've passed agent review. It then either ships them or hands them off to my coworkers for codebases where I'm not allowed to just merge PRs.

Around this workflow I had it write a small collection of skills and scripts to grease the wheels and make repetitive actions more reliable.

Since Remuda supports several terminal multiplexers, I've swapped out the backend 3 times: tmux -> herdr-tui -> herdr-core. There's experimental support for Rex but I'm waiting for it to mature a bit before I try and daily drive it.

It's not nearly as polished as some of the software factory setups and fancy GUI orchestrators I'm seeing, but it's light, hackable, and so far the bitter lesson has kept this working really well for me so I haven't felt much of a need to try anything else.

I have been trying out Muse for a while but it hasn't found a solid place in my life yet. Still too unreliable and too many privacy concerns to justify giving it a lot of access.

I use pi as a harness, and Paseo as a Mac UI (because I need a Ui).

I tried a bunch of models, but so far deepseek flash from October is just dirt cheap; and roughly on par with the frontier open weight models.

It’s still somewhat behind closed weight models (I only know Gemini Argon) in that it produces lots of text that is approximately correct —- frontier models seem to be very good at producing concise text that’s very correct.

I pay per token, using an EU hosting provider (melious.ai).

But I can fill in the gaps myself, so for me developing for fun, this is enough.