Here's my personal workflow. I've built a relatively small (<5k LoC Go) CLI to facilitate it. Mostly vibe coded in a dogfooded fashion.

When I want to start an agent, I run a small command that takes in a prompt (and some optional flags but mostly the defaults are what I want). It does this: 1. Check out a worktree from a repo (defaults to my bread and butter repo with completions for the most active ~5 repos in my org) 2. Craft a command to launch an agent (supports codex, opencode, and gemini-cli because my company only gives us OpenAI keys). 3. Launch the command in a tmux session. 4. Exit 0.

There are some light integrations with Jira, GitHub, etc but they're accomplished by shelling out to other tools.

If I want to manage sessions, I've got tmux + fzf + other unixy tools old & new to manage them.

If I want to manage sessions remotely, there is vibetunnel. It would be cool to have a slicker AFK experience, but it works.

My silly little piece of vibe coded slop and duct tape, for my use case, is largely competitive with most of the offerings on the market, outside of the ones that do cloud-based environments. Some of these projects are VC funded teams of people working full time, I'm assuming. What a time to be alive.

I'm also amazed at how readily projects like this just embrace Claude Code lock-in. Is there really anything specific about it that other agent harnesses don't support? I haven't used it yet, but so far it just seems like it benefits in mindshare alone from being the default/first mover, not because it supports any particular feature the others do not. TBF I do hear that it is quite good.

Quality & polish is a compelling reason to _use_ something, but it's not a compelling reason to build walls IMO, especially in a space like this where context engineering techniques, prompts, etc are in no way secret sauce and can be readily copied.