HhiveyBLOG

How to Coordinate AI Agents Like a Manager, Not a Tab Juggler

You have five terminal tabs open. One agent is refactoring auth, one is writing tests, one finished an hour ago and you haven't looked at it, and one is waiting on a question you forgot it asked. You think you're coding. You're actually trying to coordinate AI agents from memory, and memory is the worst place to keep that job.

Addy Osmani put it plainly: working with a coding agent all day is management work. That means picking tasks, delegating, evaluating, and deciding what ships. Once you see it that way, the fix gets obvious. You don't need to manage harder. You need to hand most of those duties to roles, and keep only the one that needs you.

I'm Alyssa, and I build hivey. This post is the method I use, and it works even if you never install anything.

Coordinating AI agents is management work

Here's what goes wrong when one person holds all four duties for every tab:

The problem isn't the agents. It's that you are the only connection between them. Every handoff, every status check and every "wait, did you finish?" goes through you.

A human team lead in that position would add structure: a written plan, someone who assigns work, a reviewer who isn't the author, and a clear sign-off step. Your agents need the same thing.

The four duties, mapped to roles

The trick is to stop thinking of duties as things you do and start thinking of them as roles someone fills. Three of the four can be filled by agents. One stays with you.

DutyWho should hold itWhat it looks like in practice
Picking tasksA written plan the team can readA short file of goals, open tasks and decisions, kept up to date
DelegatingA coordinator agentOne agent that splits the goal and assigns work to the others
EvaluatingA reviewer that didn't write the codeA second agent, ideally a different model, that checks the work
Deciding what shipsYouA clear yes or no on finished, reviewed work

The reviewer row matters most. An agent checking its own work tends to agree with itself. A separate agent with a different brief ("find what's wrong with this") catches things the author never looks for. If you can run a second model, like Codex reviewing Claude Code's output, even better: they miss different things.

A checklist to coordinate AI agents without any tool

You can do this today with plain terminals and a text file.

  1. Write the goal once. One paragraph, in a file every agent can read. Stop retyping context per tab.
  2. Name one coordinator. Pick one session whose only job is to break the goal into tasks and hand them out. It doesn't write code.
  3. Give each worker a narrow role. "Tests for the billing module" beats "help with billing".
  4. Route messages through the coordinator. Workers report to it, not to you. You read one summary instead of five scrollbacks.
  5. Add a reviewer that isn't the author. Every finished task gets a second pass before it reaches you.
  6. Keep decisions in a file. When you choose an approach, write it down so the next task (and the next session) doesn't argue it again.
  7. Reserve your time for sign-off. Your job is the final yes or no, with the review notes in front of you.

Done by hand, step 4 is where it breaks down. Copying messages between tabs is exactly the babysitting you wanted to stop. That's the part worth automating.

How hivey coordinates your agents

Hivey is a terminal workspace and coordinating agent for swarms of AI agents. You talk to Hivey, and it creates and manages your swarms. Each swarm gets its own workspace and a dedicated coordinator that directs its agents, keeps the work moving, and reports back to you. Teams can use Claude Code and Codex in different roles.

Start it where your work lives, and check what's already running before you create anything new:

hivey
hivey session list
hivey swarm list

Then give it the goal and let it put the team together:

hivey swarm new "Add usage-based billing, with tests and an independent code review"

The coordinator assigns work to agents in separate terminal spaces. Agents message teammates directly, with queued delivery, inboxes, and message history, so the handoffs from step 4 no longer go through you.

In the commands below, replace <swarm> with your swarm's name from hivey swarm list. When you want to add context or change direction, talk to the coordinator instead of each worker:

hivey msg send <swarm>/coordinator "<message>"
hivey msg log --swarm <swarm>

To see where things stand without reading every pane:

hivey swarm info <swarm>

For the "decisions in a file" step, a swarm can optionally keep a local Markdown wiki for sources, findings, decisions, and reusable knowledge. You can open the vault in Obsidian, and it lives in files you and your agents can inspect.

If you want the coordinator to revisit progress on its own, add a scheduled check-in:

hivey swarm schedule add <swarm> --every 2h "<instruction>"

Need to hold things while you think? hivey swarm pause <swarm> holds message delivery, and hivey swarm resume <swarm> releases the queue.

Keep the ship decision, even away from your desk

The one duty you keep is deciding what ships. That shouldn't chain you to your chair.

Detach with ⌥Q (or ctrl+b, then q) and the agents keep running on your machine. Run hivey to reattach later. Hivey runs locally, not as a hosted service, so your Mac or Linux box has to stay awake and connected while the work continues.

If you connect Slack, you can talk to Hivey in its channel and to each swarm's coordinator in that swarm's channel. The coordinator reports back on the work. You read the summary and say yes or no. That's the management job reduced to the part that needs your judgment.

For late nights, hivey quiet on silences sounds and holds Slack posts while agents keep working. When you turn it off, the held posts arrive as one catch-up message.

Try it

On an Apple Silicon Mac:

npm install -g @hivey/cli
hivey setup

On Linux or an Intel Mac, build from source from the GitHub repo. Setup, commands, and everyday use are in the docs.