1) Subscriptions: Claude Ultra for coding and analysis (Opus 5.5 mostly, sometimes Fable for more complex questions) + ChatGPT Business (premium seat) for more allround tasks, although I now mix in Codex (Astra) too for reviewing work or if I hit a usage limit.

Also I have a Cursor Pro plan but I notice I'm not using that enough actually. Only for backup and for file-level changes.

2) Others: I have a DGX Spark running Qwen 3.8 27b (for confidential/sensitive work, narrow coding and narrow agentic work), but it's often running ComfyUI instead, with models like moss-image/minimax h3 for image/video asset generation.

I also use various APIs incl. OpenRouter, but only for production, user-facing agents.

3) Harness: Claude, sometimes Codex

I tried OpenCode, did not like it the CLI nor the results, as compared to Claude. I want to but haven't really tried Pi yet.

4) Agent orchestration: only custom tooling now, heavy user of Mastra for custom agentic workflows incl. production agents. Fantastic open source framework (not affiliated).

5) Agent platforms: Tried OpenClaw earlier this year, when it was such a hype. Did not like it. Every update something broke, and I spent more time fixing things than I got value out of it. Switched to Hermes Agent running on a Mac Mini (I know...), and I'm still using that for various Hermes agents for all kinds of agentic tasks on the go, and so is my wife.

#1 use case is capturing videos, podcasts, and posts online that I like, or meeting recordings, transcribing them with whisper, organizing them on my NAS, ... It will also distill skills from useful info I indicate, which are then pushed to my skills collection git repo.

#2 use case is website maintenance (pushing new blog posts etc..) for a family member.

#3 delegating work to Claude for coding on the go. I prefer to do this via my Hermes agent rather than the Claude app, as it has the full local context & tooling (e.g. iOS simulator, screen recording the end result and sending it to me when done).