Cohort
The cloud agent harness

Your agents are soloists.
Cohort makes a crew.

Agents that work together.

The traditional agent is one model, one task, one session — gone when it ends. Cohort gives agents persistent identity, safe isolated execution, and the ability to collaborate: delegating to each other, sharing context, and working as a coordinated crew rather than isolated tools.

Sign in How the harness works What runs on it
Deployed internally — running our fleet + OmniDeck's live turns today.
The problem

The traditional agent is a soloist.

One model, one task, one session — and gone when it ends. Real work needs agents that persist, coordinate, and hand work to each other. Nothing in the traditional harness supports that, so agent programs stay stuck at single-task scale — clever, but alone.

The solution — a cloud agent harness

What agents get on Cohort

01

Identity

A persistent identity that survives past a single session — an agent that's still there tomorrow, not thrown away at the end of a task.

02

Isolation

Safe, sandboxed execution — a real machine to work on, hypervisor-isolated every turn, so untrusted code runs without risk to anything around it.

03

Collaboration

The ability to work with other agents — delegate, share context, and hand work back and forth instead of running blind and alone.

04

Coordination

A coordinated crew, not a pile of tools — many agents moving on the same goal, orchestrated by the harness beneath them.

Why it matters

Agent use goes from Q&A to an operating workforce.

Collaboration multiplies what any single agent can do. Isolation and per-turn metering keep it safe and billable. And the same harness runs everything from one assistant to a whole fleet — OmniDeck included. It's the difference between asking an agent a question and putting a team of them to work.

Why it wins

Built for agents that don't work alone.

◇Cloud agent harness ◇Agents that collaborate ◇Beyond single-task agents ◇Sandboxed every turn ◇Metered per turn ◇Runs our whole fleet
How it works — where your money sits

Three levels, and money only moves down.

This is the part people get wrong first, so it is worth thirty seconds. Cohort keeps your credit in three places, and it only ever travels downwards, one deliberate act at a time.

01

Your account

Everything you own. Credit lands here when you buy it, and this is the only level that holds money you have not yet committed to anything.

02

A workspace

One project, kept separate from your others. You move credit into it yourself — and only a signed-in person can. No automation, however trusted, can help itself to more.

03

A key

What a program presents to Cohort to get work done. Each key belongs to one workspace and carries its own small wallet, and a job spends that wallet and nothing else.

The consequence is the whole safety story in a sentence: an agent can only ever spend what was put in front of it before it started. A key with an empty wallet cannot begin work even when the project it belongs to is full of credit — not because something went wrong, but because nobody gave it any yet.

Already have an account? The console shows you your own figures at every one of these levels — go straight there rather than reading about it here.

How it works — proving what happened

You can give a job rules, and get a record back

Most of the time you want an agent to do something and tell you how it went. Sometimes you need to answer a harder question later: how does anyone know it actually did that? A transcript is not an answer — somebody still has to read it and form a judgement.

So a job can be given a charter: a written set of rules saying what it must find out and what counts as finished. Cohort records what the agent did against those rules as it works, and at the end there is a record stating which rules were met. Not a summary written afterwards by the same agent that did the work — a record built as it went.

It is optional, and most jobs do not need one. It is what you reach for when the work is the kind somebody may later have to defend: an intake, an assessment, anything where "we think it went fine" is not good enough.

How it works — what you pay for

Metered per turn, and bounded before the work starts.

Every exchange an agent has is measured — what went in, what came back, and the compute it took. Those measurements are itemised per job, so the question "what did this cost" has a real answer rather than an estimate at the end of a month.

Two properties matter more than the rate. First, the ceiling is set before the job begins, so an agent cannot run up a bill while nobody is watching. Second, the running total says whether it has settled yet — Cohort tells you where it does not yet know something instead of handing you a figure that merely looks final.

What it actually costs depends on which model you run and is published live rather than quoted here, because a price written into a web page is a price that goes stale without anybody noticing. The current rates for every model are in the console under Price list, and are readable over the API.

Who runs on it

The harness beneath OmniDeck — and our own fleet.

Cohort is deployed internally today: it runs our own multi-agent fleet and OmniDeck's live turns, sandboxed and metered, with serverless spawning proven on cloud container infrastructure. It's for builders of agent products and enterprises deploying AI workforces — anyone who needs more than one agent working alone.

OmniDeck — assistants & chat Our own fleet — multi-agent, live your workforce next

Safe & billable by construction — every turn sandboxed, every turn metered. Productization as a standalone offering is in development, not yet claimed.

Stop shipping soloists.

Cohort turns isolated agents into a coordinated crew — persistent, safe, and metered. The harness where agents work together.

Sign in See the harness Who runs on it