Rough stream of thoughts... no AI was used in the process

Managing agents in shifts

The biggest shift in my productivity as a founder juggling engineering, sales and management is starting to use my Linux box at home (a VPS would also work) and run agents remotely. Having a box that is always on allows me to think about managing agents similar to managing people.

At the crux of this shift is muxflow. It’s a product I built for myself but it was so transformational for me that I decided to open source it. Muxflow allows me to easily connect to and work on any machine with simple SSH and run agents across these machines seamlessly. Everything works as if it were on my personal device - copy and pasting, transferring files between machines, agent notifications, native git client and file explorer etc.

And it’s all cross-platform (Linux, Mac, Android and iOS) based on tmux so I don’t need to keep anything open for my agents to keep working.

My workflow is simple. I work with my agents in shifts and I think about my work as managing agents. The first shift of the day is actually the day before. At the end of every day, around 10PM - I prepare my todo list for tomorrow. That’s a practice I’ve always done as I find the best ideas and clarity come to me at night. Then I look at my todo list and ask “Which tasks can I do with agents and how do I allow them to fully complete them?”

I then kick off an agent per task. I work with agents in parallel (muxflow makes it easy) and my goal is to formulate a plan with the agent that allows the agent to fully complete the task. I identify any blockers, the expected result, the level of QA I expect from the agent and transfer all of the context to the agent. I ask the agent to save all plans and context to files. I have a git repo called ai-projects with subfolders per topic. So if I work with an agent on support for one of our customers, everything is saved to git so the next agent has all context.

Once I’m done planning I let the agent run.

From there I work in shifts. I’ll wake up in the morning, and before I head to the office I’ll review the agent work, provide feedback and kick off subsequent agents. My day will usually be full of meetings, firefighting and working with my team. While I’m in meetings, the agents work. After lunch, I’ll do another session and proceed with day-to-day stuff, then kick off another shift before leaving the office that I’ll review at night and start a new cycle.

With the right infrastructure, I find that agents can fully complete 90% of my hands-on tasks, usually by the next shift.

Beyond muxflow, I have a few skills and processes that help me empower my agents to work end-to-end:

  1. browser-skill - A wrapper above browser-use CLI that provides browser isolation and teardown between agents and import/export cookies across SSH. The wrapper allows me to easily transfer cookies to my Linux machine, and my agents to work on several browsers in parallel without stepping on each other’s toes.
  2. gdev_cycle - A skill for development that instructs my agent to run independent review cycles and QA. It’s a simple prompt but using it brings the PRs my agent works on 3x further on their own. You can review the skill here.
  3. gdev_subagents - A skill that instructs claude/codex to plan big implementations in waves of gdev_cycles. I use it for very large implementations that require multiple agents.
  4. bro - Tells the agent to repeat the last message in simple terms.
  5. cua-driver - I don’t use it as often as the browser skill, but every once in a while I might need the agent to use my computer.

← All posts