Hey {{first name | there}}. Over the course of a year, I have gone from "nah, I don't think agents are that cool" to "wow, this might actually be useful".

A lot of that change came from how I run them, not which model I use. Central to how I use agents is a multiplexer named herdr.

In today's technical notes:  

  • Why I am no longer on tmux

  • Why I think you should use multiplexers

  • Checking on agents from my phone

📰Technical notes: How to get more out of your agents

Why I am no longer on tmux

I used tmux for years. My config was small: tmux-powerline for the status bar, tmux-pain-control for splits, tmux-resurrect to bring sessions back after a restart, and a few bindings to resize panes.

That setup was built for shells and editors. Once most of my panes had a coding agent in them, two things kept getting in the way.

The first was knowing which agent needed me. tmux does not know whether the agent in a pane is waiting on you. If I had five agents going, I found out which one was waiting on an approval by flipping through every pane.

The second was restarts. tmux-resurrect brings the layout back, but the conversations in each pane were gone, and I had to find the session ID for each agent and resume it by hand.

The friendship-ended meme, with tmux swapped for herdr.

herdr is a terminal multiplexer built for coding agents. It keeps the tmux model of sessions, tabs, panes, and detach, and adds a sidebar that knows which agents are running and what state each one is in: working, blocked, done, or idle. It detects Claude Code, Codex, and a long list of other agents without extra setup.

Switching cost me very little. herdr's prefix is Ctrl+B like tmux, and I kept the gruvbox theme I had before. This is most of my config:

[theme]
name = "gruvbox"

[keys]
prefix = "ctrl+b"
split_horizontal = "prefix+minus"

[ui]
show_agent_labels_on_pane_borders = true

[ui.toast]
delivery = "system"

[session]
resume_agents_on_restore = true

resume_agents_on_restore fixes the restart problem. When the herdr server comes back, it restarts each agent with its own resume command, claude --resume <id> or codex resume <id>, so the conversations pick up where they stopped.

Toasts go to macOS notifications, so I hear about a blocked agent even when the terminal is behind another window.

Why I think you should use multiplexers

An agent is a long-running process. It might take forty minutes on a task, and it should not die because you closed a terminal window. A multiplexer keeps the process in a server that outlives your terminal, so you can detach, close the lid, and attach again later.

The other reason is layout. I split work by job, not by repo.

My default workspace has a tab per kind of work. explorer has one agent I use to poke around a codebase or a question. research has another for reading and summarising.

blog has Codex drafting on the left and Claude reviewing on the right, so one drafts and the other works on titles and edits.

Anything bigger, like an operator I am building, gets its own workspace.

A herdr blog tab with Codex outlining a post on the left, Claude suggesting titles on the right, and every agent listed in the sidebar.

The sidebar is the part I look at most. The red dot on demo-api is an agent blocked on an approval, and I can see it from whatever tab I am in.

These screenshots come from a separate herdr session I set up for this issue, with the same layout and theme as mine and the project names changed.

Checking on agents from my phone

herdr has no mobile app. When the terminal is 64 columns wide or less, it switches to a single-column layout that fits a phone screen.

My Mac and my phone are both on my Tailscale network, so I can SSH into the Mac from anywhere without opening a port to the internet. From SSH app on my phone:

ssh [you]@[mac-name]

herdr

That attaches to the same session that is running on my desk. If an agent is blocked on a command, I can read what it wants to run and approve it from the phone.

herdr on a phone over SSH, showing an agent waiting for approval to run go test.

Most of my agent work now runs this way. I start the long tasks at my desk, walk away, and check in from the phone when one of them needs me

🌍IN THE ECOSYSTEM

How to work with Herdr covers the three ways to run herdr: locally, inside an SSH session, and with herdr --remote, including the section on working from your phone.

Tailscale SSH walks through letting Tailscale handle SSH authentication between devices on your tailnet, so you are not managing keys for every machine.

My AI Adoption Journey by Mitchell Hashimoto. He goes from chatbots to running agents at the end of the day and delegating tasks while he works on something else, and explains why he still only runs one background agent at a time.

⏱️UNTIL NEXT TIME

If you run more than one agent, put them in a multiplexer this week and give each kind of work its own tab.

I'd like to hear how you run yours. Are you on tmux, zellij, herdr, or a stack of terminal windows? Reply, and I will feature a few setups next time.

Know an engineer who will find this helpful? Share this link with them

Jubril Oyetunji
CTO, EverythingDevOps

HOW DID WE DO?

Login or Subscribe to participate