How to Better Manage Your Agent Workforce

A few Mac minis running coding agents can keep a lot of work moving.

They can also leave you wondering what is actually getting done.

An agent finishes a fix while another gets stuck. Somewhere in a terminal, a bug has been found that nobody else knows about. You spend the next hour checking sessions and copying updates around.

At that point, you have become the messenger for your own bot army.

The way out is to manage your agents like a team. Give the work a home before someone starts it. Make sure each agent has the access it needs. Expect a useful handoff when the assignment is finished.

That applies whether you have Mac minis on a shelf or agents running on other hardware you already own.

You can hack it together. Then you have to keep it working.

You could put a board in Notion and connect your agents through MCP. Add scripts to move the work around and hope every agent remembers to report back.

It works until an update goes missing. Then you are opening terminals again to figure out what happened.

Notion can hold the work and let assistants update it. That does not make your homemade setup a reliable way to run a fleet. The part that connects a card to an actual running agent is still yours to maintain. Notion MCP.

Notion was not built for this job. Hell, half the time it feels like Notion does not know what it wants to be.

There is a better way: One Horizon.

It gives you a framework for connecting agents to the work, with a kanban board where you can follow their assignments. New discoveries can go into triage until you decide what deserves attention.

And you can use it for free. One Horizon pricing.

Add an agent without rebuilding the setup

For hardware you manage from the terminal, I would start with the One Horizon CLI.

Install it on the machine and sign in:

npm install -g @onehorizon/clione auth loginone ws

From the repository the agent should work in, create the agent and check its setup:

one agent createone agent doctorone agent startone agent status

The machine needs a working development environment. The CLI requires Node.js 20 or newer and Git. Resolve anything doctor flags, including authentication for your chosen runtime, before starting work. The terminal guide covers the details.

You can use the desktop app for guided setup instead. For a shelf of Mac minis, the CLI would be my starting point.

You are also free to change which agent does the work. You might start with Claude Code and add Cursor for a different assignment. Custom agents can connect through ACP, provided they implement the supported contract. The framework is also being extended to external A2A agents. Custom ACP agents.

The point is to keep your way of working as you add agents. Trying a different agent should not mean inventing a different process.

Stop connecting the same account on every machine

Say your agents need documents from Google Drive.

With a setup you have pieced together yourself, you might connect the account on each machine. Add another Mac and you repeat the process. Later, one connection needs you to sign in again, and you have to find out which machine is affected.

With One Horizon's managed connections, you handle that access in one place. Connect Google Drive once and make it available to the managed agents that need it. A new worker uses the access configured for its work.

You do not have to paste credentials into prompts or manually copy them between computers.

If the connection needs attention, you deal with it centrally. An expired or revoked connection should not send you around the room signing in on every Mac.

Skills follow the same approach. You manage the skills an agent should use through One Horizon, rather than maintaining a separate installation on each machine.

This is part of managing a team. A new team member needs the right access before they can contribute. An agent does too.

Give the team a board to work from

Imagine a customer reports that an export only includes the first page of results.

Create a work item in One Horizon. Explain how to reproduce the problem and what a correct export should do. Keep the assignment focused on pagination.

Then delegate it to an agent.

That is easier to manage than opening a session and typing "have a look at exports." The task already has a place on your kanban board. You can come back to it without remembering which terminal you used.

You can also put it through a workflow so another agent reviews the implementation. One Horizon lets you assign agents to different steps and pause for human review where needed. One Horizon workflows.

If the review runs on another machine, that machine needs access to the code being reviewed. Use an accessible remote for the branch handoff. Worktree isolation gives separate tasks their own working copies. Local agent setup.

The board becomes your starting point for understanding the work. Open a task to see what came back. Look at what is waiting for review before sending out another assignment.

Make the result useful to whoever comes next

Delegated agents report progress and outcomes through their recorded runs. MCP gives agents another way to update the work itself, including when you start a session outside a One Horizon workflow. Local agent reporting.

A useful instruction is:

When you finish, update this One Horizon task. Explain what changed and link to the result. Say what you checked. If you are blocked, explain what you need.

For the export example, "done" is not enough. The report should point to the fix and say whether the pagination test passed. If the agent could not run the test, it should say that.

Now suppose it also notices that dates near midnight appear on the wrong day.

That should become a follow-up in triage. Have the agent check for an existing issue before creating one through MCP. Include a failing example and link back to the original task. One Horizon's MCP tools let assistants create work and update existing records.

You can decide whether the date bug should move onto the board. Another agent can then take it on without repeating the investigation.

Creating a triage item does not start an agent by itself. It still needs to be assigned for execution. Where an agent handoff is configured, One Horizon supports delegated sessions for another eligible agent to claim. Agent delegation.

That gives discoveries somewhere to go without letting every agent quietly expand its own assignment.

More hardware should mean less chasing

Once this works, you can start your day at the board and review the export fix. The date bug is waiting in triage for you to decide whether it should go next.

You are managing work instead of collecting updates from machines.

Keep the meaning of finished clear. A successful agent session does not mean the change has been approved or shipped. Its output still needs review. Session outcomes.

If the review column keeps growing, adding another Mac mini may not help yet. Finish the work waiting on you before filling the queue further.

Start with one agent taking a bounded task through review. Get its access right and make its follow-ups useful.

Then add more agents to a team you already know how to manage.