Frontiers models are getting smarter every week and models like Sol, Opus and Fable can take on much bigger things than fixing bugs and cleaning your inbox. You just need a system that can put them to work.
This week we’re opening the beta of our agent manager in Tability.
- Start with a goal
- Assign it to the agent manager
- The agent manager will break it down into an execution plan, recruit an AI team, and put them to work until your goal is achieved.
I’m talking about something like “increase our traffic value from $15k to $20k”. A goal that can’t be resolved in a single session. Something that you’ll need to come back to for weeks, no matter how much output you can get out in a day.
It takes 2 minutes to deploy a new team, and you can scale from 0 to 50 agents working on growth in 20mins.
The video below shows the full setup in 15mins. Yes, it’s more than 2mins, but it’s because I’m talking through everything with my french accent.
Timeline
- 0:00-2:45: background concept
- 2:45-3:30: assigning an AI manager to a goal
- 3:30-5:30: heartbeat creation in Codex
- 5:30-8:25: AI manager breaks goal into execution plan
- 8:25-9:45: Reviewing and approving the proposed AI team
- 9:45-11:50: AI team creation (by the AI manager)
- 11:50-end: me selling you on all the great things you’ll be able to do because I can’t stop myself from talking about it.
You can try this today at tability.io. Just connect the Tability MCP Server to Claude or Codex and execute the steps in the video (personally I prefer Codex because I find it better at running uninterrupted).
Why you recruit managers, not agents
If you try to scale agents by creating them yourself, you become the bottleneck. Ten agents is fine. A hundred is a full-time job of writing prompts, defining roles and remembering who is meant to be doing what.
So the model is top-down. You start with a goal, and instead of handing it to a team of humans, you hand it to an AI manager. The manager reads the goal, breaks it into smaller pieces, and recruits its own team of specialists to work on those pieces.
You manage managers. They manage the rest.
The other half of this is that we are not replacing your existing setup. The manager and its team run on your Claude or Codex account, with the tools you have already connected there. Tability holds the business context – the goals, the plans, the progress, the history – and your agent environment does the work.
The setup takes about two minutes
You need a Tability workspace connected to whatever you are running, Claude or Codex.
In the demo I have a plan with a goal on it: increase the Ahrefs traffic value of a site from $15k to $20k. Instead of assigning it to a human, I assign it to a manager and pick the host. Tability generates an instruction. I copy it, paste it into Codex, and run it.
From there Codex sets up a heartbeat. It asks which model to use and how often the manager should run. I went with every 30 minutes.
You can push that further. Every 20 minutes, every five minutes, 24/7 if you want. The only real limit is your ability to keep up with what comes back. Every run produces updates, proposals and questions, and there is no point generating more of that than you can read. Twenty to thirty minutes is a good starting point.
First run: a plan and a proposed team
Once the heartbeat exists, the manager moves into planning. In the video I force a run rather than waiting, but this is exactly what would happen on its own.
The manager comes back with two things.
First, a sub plan. The goal has been broken into key results – repairing broken pages, lifting traffic value on commercial pages, and so on. All the execution for that goal now lives in a structured plan rather than scattered through a chat log.
Second, a proposed team. In this case an SEO optimiser, a repair specialist and a digital PR specialist, with roles written for this specific goal rather than pulled from a template.
You do not have to go looking for any of this. It lands in the input required section on your homepage. You review the proposed team and approve it, and the next heartbeat creates the agents, each with its own dedicated chat and its own slice of the plan attached.
Set this up in the morning, come back at lunch, approve a team. That is the whole first day.
They will ask you for access
Here is the part I want to be honest about. Early on, you will get a queue of agents asking for things. CMS access. Repository access. Whatever they need to own their job end to end.
That is not a bug in the process, it is the process. The agent tells you what it would need to work without you in the loop, and you decide. You can enable it, or you can tell the manager to work around the constraint and it will adjust.
The easiest way to handle it is in the original manager thread, so the context stays in one place rather than being re-explained to every agent.
After the first few days this settles down, and the agents run without asking you for much at all.
What you get that a thread does not give you
The real difference shows up a week or two in.
A goal like traffic value does not move because of what happened today. It moves over weeks, and the only way to know whether the work is compounding is to track it properly. With this setup you get the execution plan, week by week progress on each key result, charts, and updates written by the manager explaining what changed and why.
You can still drop back into the Codex or Claude session to see exactly what an agent did. But you no longer have to, and that is the point. You can tell which agents are doing good work and which are not, from a dashboard, without reading anything.
And if you want to steer, you reply in the comments. The manager picks it up on the next heartbeat.
Internally we are running about 25 agents on this, 20 of them under my management, and we are still scaling the team up. The point is not the number. It is that adding the twenty-first agent costs me nothing, because I did not create the first twenty either.
Try it today
Start a trial at tability.io, connect your workspace to Claude or Codex, and assign a manager to one goal. Assign manager, submit, paste the prompt into your tool. Everything else runs from there.
You can speedrun it like I did in the video by forcing the heartbeats, and it is worth doing once to see the shape of it. But the better version is the boring one. You set it up before bed, close the laptop, and there is a plan, a team and a pile of feedback waiting for you at breakfast.
.png)
.jpg)
%20(1).png)
