Quote :

"what IDEs/harnesses y'all using?"

Claude code for my work daily driving as the basic primitive, super customized with hooks and some clis + skills to match my workflow. I multiplex claude codes through an app called subspace -https://subspace.build. It works with codex, Gemini cli, opencode, pi… basically any cli based agent you might want to use can be invoked. It has built in workspaces and cross-agent memory. I sound like a salesman, but it’s almost peerless… I feel like I’m hamstrung when I’m not using it because it’s such a good piece of kit. I tried opencode as the basic primitive with this stack, and it works, but I just prefer Claude code as the base harness. I’ve been meaning to try pi, but have been busy with other shit. I also run Cursor, for when I need to do quick file edits, but for the most part I don’t really even look at code anymore. I pretty much stick to Opus/Sonnet for day to day stuff and have the Max personal plan + api billing at work.

On the side I’m building my workflow out into something more hands-off via my Hermes agent. Essentially trying to replicate my day to day harness setup from above, but with multiple layers of agent-owned orchestration to pick up work, reason over it if needed, build it out, and submit a PR that I look at and give feedback on for iteration. Think “go look at the tickets under Milestone XYZ in Project ABC in Linear. LMK if you have questions, otherwise bang it out using [individual||stacked] PRs” - here I’m using deepseek 4 flash almost exclusively ATM.

[Edited on May 23, 2026 at 5:02 PM. Reason : Page 3]

5/23/2026 5:01:16 PM

A fun goal I have with Hermes, soonish, is to get it a voice agent and do whisper/TTS so I can voice chat with it, either via discord or something else, because every remote agent solution I’ve seen so far sucks huge fucking asshole, especially when you need to approve requests.

5/23/2026 5:06:29 PM

ClaudeCode for personal and OpenCode for work. Did GeminiCLI last year because we have a deal with google but agentic stuff was getting caught in loops and tool use sucked.

^ I did an iOS app to nanoclaw with speaking. Could do discord audio too and telegram. I did a deep dive on audio. Best performance is the OpenAI voice model but it's behind the times and it can't do tool use. I used pieces of this architecture: https://www.ntik.me/posts/voice-agent and got latency down to a couple of seconds. Perfectly acceptable for common tasks but not very conversational.

5/23/2026 11:08:47 PM

^ Nice.

Apparently there's a 1p Google Meet plugin for Hermes... gonna try that sometime to see what it does out of the box. We use 11Labs at work and it does pretty decent conversationally, so I may stick with that, or I may try out the new model from Mira Murati's lab just to fuck around.

5/24/2026 9:25:24 AM