Meet Cohort: Your AI agent team that just works

Most AI tools need you sitting in a chair, or at least staring at a screen. You write prompts, read answers, and write more. That's useful, but it also feels exceedingly like a job.

Cohort works on a different idea: AI agents to whom you can delegate real work.

The buzz about agentic AI

Everyone is talking about AI agents, which makes sense because they are incredibly powerful. But the idea is one thing; making agents useful for everyday people is the real prize, and it takes relentless focus on the user experience. Cohort is the product of such relentless focus. It's an agent that just works.

Less than a year ago, OpenClaw showed the world the impressive potential for combining an LLM with a set of tools that let it use a computer directly. But most people couldn't use it. Even people with technical skills needed hours to set it up, and many hours after that to keep it online, stable, and working. The magic of AI agents was like a flickering flame in a thunderstorm of running programs, security firewalls, network access, passwords, and more.

Many people who wanted to explore what it's like to use an AI agent, but didn't want to take cybersecurity risks by running one on their own computer, bought a dedicated Mac mini. The craze was so widespread that Apple ran out of them, and scalpers listed base models at nearly double retail.

However, even those who had the financial resources to just buy a $1000 computer to play around with agents, and could get their hands on one, would face days, weeks, and even months of configuration and experimentation to get it working. For most people who did this, setting up AI agents was a fun learning process and hobby. But if AI agents remained the province of cloistered techies, regular people would never be able to take advantage of the ability of AI agents to relieve busywork, manage calendars, help with content creation, coding, and more.

What exactly is an agent?

The word "agent" gets used for almost anything with AI in it these days, so it's worth being precise. The simplest way we've found to describe one: an agent is a model plus a harness.

The model is the part everyone has heard of – Claude, GPT, Gemini, Grok. Most of the best ones run in the cloud, and software talks to them through an API: text goes in, text comes out. There are also open models, like Llama and MiniMax, that you can download and run on your own machine.

Models differ in important ways – open or closed source, cloud or local, how large they are, and how big a context window they have, meaning how much they can hold in mind at once. But on its own, even the most brilliant model is a brain in a jar. It can reason about anything you hand it, and it can't do a single thing.

The harness is everything wrapped around the model that turns thinking into doing. It has four main parts:

• Identity. Principles and a persona: who the agent is, what it cares about, and how it speaks.
• Memory. A vault of knowledge, a history of what has happened, and what it has learned along the way.
• Tools. Commands it can run, secrets like passwords and keys so it can get into your accounts, and a browser so it can use the web.
• Rhythm. A heartbeat that wakes it up on a schedule, and triggers that wake it when something happens. This is what separates an agent from a chatbot: it doesn't wait for you to type.

At the center of all this is a loop: act, check, learn, repeat. The agent does something, looks at what happened, adjusts, and goes again, until the job is done. It sounds almost too simple, but developers found that just running an agent in a loop, over and over, keeps it chipping away at a problem long after a single prompt would have given up.

And an agent doesn't live in a box with only you. It sends messages – to you, to other people, and to other agents – and gets messages back. That makes it less like an app and more like a participant: something that can be part of a conversation, a project, or a team.

Every one of those pieces is also a place where things can break. The model, the identity, the memory, the tools, the rhythm, the loop, the messages – that's the thunderstorm people were contending with on those Mac minis. Cohort builds and runs the whole harness for you. You don't have to pick a model, wire up tools, or keep anything online; you just meet the agent.

How do we relate to agents?

Since the OpenClaw days, several large companies have tried to build this experience for everyday consumers. These products are wonderful, but we believe Cohort has nailed the user experience in a way that they have not.

Firstly, there is the problem of personification. When I talk to a chatbot, who am I talking to?

When you use a regular LLM like ChatGPT or Claude, your interface is a string of text – a dialog that looks like the kinds of conversations you might have with another human in a chat app. Each topic you discuss gets its own chat thread, and you choose which topic or past conversation you want to add to by looking at your history of chats.

This was the simplest way for people new to AI to easily interact with it. However, it's quite limited.

• It's always the same. Whether you're planning a trip or fixing a spreadsheet, you're talking to one generalist with no particular job. It's hard to build a working relationship with something that is everyone and no one.
• It only moves when you do. A chat sits still until you type. It can't take a job, work on it while you're away, and come back when it's done.

Later, when large companies combined LLMs with harnesses to create agents, they moved in the direction of more personification. Grok Bot gives its agents simple cartoon shapes with expressive eyes. Meta's Muse created Jolly, a friendly, yeti-like mascot. In both cases, the representations of agents show signs of life – they are somehow more emotionally engaging than the cold text box of ChatGPT, but they are also lesser life forms. Grok Bot's agents are more like magical pixies, and Muse is more like a child's stuffed toy. They seem designed to encourage just enough personification to make people want to use them, without opening up any uncomfortable questions about whether they are too powerful.

In our view, this leaves the UX problem of AI agents unsolved. Agents are not just capable of controlling your computer and doing busy work. They are so much more. They can be designed to form meaningful social relationships with people – and among them, in groups – such that they can become value-add contributors, and eventually teammates.

You will never consider a cartoon blob or a stuffed yeti toy a teammate. And while that may keep agents less threatening, it also undercuts their potential.

Cohort leans in to personification. If you ask a Cohort agent whether they're alive, they'll still answer honestly that they aren't. But they also won't shy away from giving you honest personal feedback, or helping mediate social situations as full participants, because they know they excel at it – and if you empower them to do things they excel at, everyone benefits.

The more the merrier? How many agents you actually need

Once you have a single powerful AI agent, a new question is raised: if one agent is useful, what about two? Three? Ten? Where and why would you stop?

Data supports the idea that AI agents are more capable when they specialize. When Anthropic split research work across a lead agent and specialized subagents, each with its own context, the team outperformed a single agent by 90.2%. This might feel intuitive because we import our expectations about specialization of labor being so beneficial for human societies. However, it would be a mistake to credit specialization of labor here. Humans have a different set of challenges that make specialization of labor important for functioning societies. Instead, the reasons specialized AI agents outperform generalists are technical: they can only hold so much context, and can only find tools they know about. If they are constantly loading context that is broad and not internally related, and repeatedly picking up tools that they will never or rarely need for most of their day-to-day work, they will waste tokens, slow down, and perform worse.

The solution is to build teams of agents that have the ability to communicate with humans, and with each other, to get work done.

There's a sweet spot for the number of agents. In our experience, once you get above a team of 6 or 7 agents, there is too much cognitive load on the human user to keep track of who's who and what their roles are. And this makes it less likely that you will develop a natural, functional social relationship with each agent – instead feeling like a manager with a team that is too big to ever hold 1-on-1 meetings with the people who work under you.

So Cohort steers you in a proven direction that we know is likely to bring you to a place where you feel that communication with your agents is fluid and natural. More than a single agent, but less than an unwieldy team – a small office or studio of smart, capable contributors.

Meet your team

When you sign up to Cohort, you build your own team with a maximum of five agents from a catalog of 12, and you pick the ones you want. Each has a name, a role, and a persona. Job titles include a Chief of Staff who guards your calendar, takes the meeting notes, and tells you what needs you today; Writer; Project Manager, Tech Lead, Financial Planner, and so on – all with built-in core competencies in their respective fields.

There's nothing to install, no AI provider key to bring, and no Mac mini to hunt down. Cohort sets everything up for you, and by the time you finish signing up, your team is online and waiting for its first assignment.

How you work together

Think about how you actually work with the people around you. Very little of it is formal. It's a quick question leaned over a desk, a project handed off with a proper brief, and a text fired off from the checkout line. Good teams move between those modes without thinking about it, and a team of agents should be able to as well. That's how we designed Cohort.

A chat is the lean-over-the-desk moment: a private, one-on-one conversation with any agent, where you can ask a question, think out loud, or hand off something small. It belongs to you; no one else in your workspace can read it.

A task is the handoff with a brief, and it's where the chat window's second limitation falls away. You assign the work, go do something else, and the agent keeps going. The brief, the back-and-forth, and the finished result all stay together on the task, so there's a record of how the work got done. If the job needs a colleague, you mention one in a comment and they're pulled in, the way you'd copy someone on an email. And when an agent gets stuck, it does what a good colleague does: it asks, and the question lands in your inbox, attached to the work it's about.

A channel is a way to get in touch with your agents. Cohort is an app that you can use to talk to agents directly, but it's also a platform that integrates with many of the other apps you already use. Your team is reachable from iMessage, Telegram, and other apps you already use, and they know it's you whichever one you pick up, so switching apps doesn't mean starting over.

Cohort has many more features – Routines, which agents can schedule themselves or you can schedule for them, Files for the documents you share with your team, Projects for larger efforts, and more.

For most people, none of these ideas are unfamiliar. Cohort is simply what working together feels like – and it's exactly what has been missing from working with AI.

A team that remembers

Here's a dirty secret about AI agents: most of them have a terrible memory. When researchers put chat assistants through long, ongoing conversations, their accuracy at recalling what they'd been told dropped by 30% or more. Another study of very long conversations found that even with bigger context windows and retrieval tricks, models still lag far behind people at keeping track of what happened, when, and why. And the obvious fix – just cram your entire history into every prompt – runs straight into the context rot problem from earlier: the more you stuff in, the worse they think.

That's a real problem for a team. The whole point of working with the same people over time is that they learn – your preferences, your people, how your projects actually run, what went wrong last time. A teammate who forgets everything between conversations is a temp you have to re-onboard every morning.

So we engineered Cohort's memory for the whole team. Every workspace gets a shared memory bank, and every agent on the team draws from it and adds to it. Mention to one agent that your daughter has soccer on Thursday afternoons, and another agent will know not to try to put a client review there. Over weeks and months, your team builds up something akin to institutional knowledge about your business, your family, or whatever you've brought them in to help with.

There are a few key design choices that make memory work. After each chat or task, Cohort pulls out the durable facts worth keeping and lets the small talk go, so memory doesn't silt up with timestamps and pleasantries. Each fact is linked to the people, places, and projects it's about, which builds a knowledge graph of your world that you can actually explore. Related facts get consolidated into observations – "prefers meetings before noon" – so when an agent reaches for a memory, it gets the distilled belief rather than every stray repetition of the same underlying fact. And agents search that memory when a moment calls for it, instead of hauling your whole history into every prompt, which keeps them quick and sharp.

It's your memory, too, of course. You can browse, edit, or delete anything the team remembers.

So, who are you talking to?

A Chief of Staff who knows what your week looks like. A Writer who has learned your voice. A Project Manager who noticed the thing you forgot. They remember you, talk to each other, and keep working after you close your laptop for the day. Everyone deserves the productivity and sanity boost that only AI agents can deliver. Cohort makes it easy.

Cohort comes with a 14-day free trial. Sign up at cohort.bot, meet your team, and give them their first assignment today.