I found a way to manage all my AI agents

Spin up enough coding agents and you stop knowing which one is where, what each one is doing, or which one is sat waiting on you. Herdr fixes that — every agent visible, all the time, in one terminal workspace.

This isn’t an install-and-demo video. It’s how I actually run multiple AI agents at once on enterprise software, day to day, at GitLab.

Herdr is a server that runs in the background. You start it by typing herdr, spin up agents or processes like you would in any terminal window, and everything keeps running — close the window, close the laptop, it doesn’t matter. If you’ve used tmux or Zellij, the foundation will feel familiar. The difference is that tmux doesn’t know or care about your agents, and herdr is built for them first.

That means status at a glance: which agents are idle, which are working, which are blocked waiting on me, which project each belongs to, and roughly what each one is doing. Agents can spawn other agents in their own full sessions — not sub-agents, real instances I can jump into and talk to myself. Each one works in its own git worktree, so nothing collides.

The part I get the most value from is the automation around merge requests. When I’m assigned a review, a herdr session spins up on its own, gathers the context, does a first pass, then dings me — I still do the human review, but I arrive informed. When a blocking MR merges, my agent cluster rebases the next cohort, waits for the pipeline to go green, and assigns the pre-selected reviewers. Working across time zones, that often means going to bed with one MR in review and waking up to it merged and the next one ready for me.

The thing I keep coming back to: working with AI is knowing when to put yourself in the loop and when to get out of the way. I could automate the full review and merge cycle tomorrow. It would also be irresponsible. For a vibe-coded side project, fine. For enterprise software, the reliability of this code matters.

I’m Sam Beckham — former Engineering Manager at GitLab, now back to being a Staff Fullstack Engineer there, working out what agentic engineering looks like in real day-to-day work rather than in demos.

In this video

• 0:00 So many agents they creep into your dreams
• 0:30 How you lose track of your agents
• 0:53 Your agents are a team that reports to you
• 1:32 Make the status come to you
• 1:57 Always-on agents vs your daily driver
• 2:25 This is herdr
• 2:42 How it works
• 3:15 Built for agents, not just terminals
• 3:19 Every agent’s status at a glance
• 3:49 Git worktrees mean agents never collide
• 4:03 Automating the first pass on merge request reviews
• 4:48 Sequencing work and unblocking the next MR
• 5:33 Going to bed with an MR in review
• 6:00 Knowing when to be in the loop
• 6:30 Wrap up