Terminal-based Agentic Programming Roundup

  • SST's opencode and OpenAI's Codex are both now viable as daily drivers for real engineering work. A month ago, the only things I could actually use were Amp Code and Claude Code. That is no longer true.

  • Google's Gemini CLI is still a bit crashy and awkward, despite the models themselves (especially Gemini 2.5-pro) working really well within opencode. Gemini CLI has a habit of going into strange infinite loops, plus it occasionally writes a shame-filled Diaryland / Livejournal entry on its own badness. Real "Uncanny Valley" stuff. Not viable for daily driver, but hella useful in gemini -p mode as a one-off CLI the other agents can delegate to.

  • Charm's Crush is a new entrant and isn't ready for prime time. But don't count them out, because they are the maestros of terminal software (I personally use their other stuff for all of my CLIs and TUIs). OpenAI should buy them.

  • Opencode used to panic/crash a lot on my Mac, but whatever it was they fixed it! If I could get it working properly on my remote dedserv (Ubuntu 20) I'd use it a lot more than I do. I'm sure I could always compile it from source, but aren't these all supposed to be DevX companies? Give me brew + apt + nix + npm options, and make it braindead simple to get going, kthx.

  • Claude Code, for some reason, is panicking a lot more than it used to. Usually when it decides to read insanely big log files despite my instructions not to. LLMs are brats. CC has been my true daily driver since mid-May, but it's slipping a bit, because....

  • GPT 5.0 in Codex is an absolute beast. The thing is just so freaking smart that, despite the actual CLI / TUI being the worst on our list, I'm using it all the time. (Claude Code Opus 4.1 plans and builds, Codex GPT 5.0 reviews/fixes. Codex also does my EM/PM work with the help of the gh and linctl CLIs now.).

  • Economic bonus points: my Chat GPT and Claude Max plans are still humming along, despite my strong intuition that I'm costing them a lot more inference money than I'm contributing in revenue. Plugging my ears and praying the rug-pull is still months away.

  • In Amp Code, GPT 5.0 is kinda wonky, and the Amp team has made comments about how they think it's not great & they recommend Sonnet still. I hope they change their tune on this!

  • Amp continues to be the best-of-breed CLI / TUI, but the economics make it not quite viable for me (i.e. they seem to be building an actual business, not a VC-subsidized loss-leadership play). Plus, they optimize for latency instead of quality, and let's be frank: Sonnet simply isn't good enough on the reasoning side for the kind of dev work I need it to do.

My coding agents are deep-thinking senior software engineers who can each work for ~60 minutes unattended while I take zoom calls. They are NOT fast interns I have to re-prompt and micromanage. I know a lot of people prefer the latter. To each their own.