Current Harness
harness for AI agents
EN · RU · ES

You have ten AI agents. You need a team.

Current Harness turns your Claude Code agents into a team. They remember yesterday, finish what they start and come to you only with what matters.

I want a team
Sound familiar?
01

Every session starts from zero

A new window doesn't remember yesterday. You explain the project and where you stopped. Again.

02

You're dispatching your own bots

The chat is buried in questions. Half are trivial, but only you can answer, so everything waits.

03

“Done” isn't done

An agent says “done” and there's nothing to check. And one of them quietly stalled yesterday.

What your day looks like

Question from the Manager

Why: two button options on the checkout page. The choice changes what buyers see.
Options:
Big button at the bottomrecommended: easier with a thumb Button next to the pricemore compact, easier to miss Decide latercomes back tomorrow morning

The whole question, with the reason and the options. One tap to answer.

Working now

Managersorting the task queueworking
Designerdrawing the lobby screenworking
Tech leadchecking the buildworking
Sound engineerwaiting for your callquestion
Game designerhanded off the thread, restingpaused
Reviewerchecking finished work liveworking

Who's doing what right now. One screen instead of ten windows.

Done: lobby screen

What's done: new screen, bigger buttons, works on phones.

verified live: opened on a phone

Accept Send back

The result, with a screenshot. You accept it or send it back.

You decide. They work.

Why it doesn't fall apart

Work outlives the session

The window closed, the context ran out. The next session picks the work up and knows what's done, what isn't and what's next.

“Done” only with proof

A number, a screenshot, a live request. “The code looks right” doesn't count.

Every agent has a reviewer

No one closes their own work. Someone one level up accepts it.

Nothing goes out without you

Releases, publishing, money, deleting data: only after your “yes”.

How it works inside
Current Harness roles and flows The human gives tasks to the manager, the manager passes them to the dispatcher, the dispatcher hands work to roles, roles bring in helpers. Cards go from the manager to the human. Thread handoff: session 1, 2, 3. Rules live apart from the engine. The log records every action. The second computer is locked behind your “yes”. People and roles Human Manager Dispatcher Role Role Role helper helper helper card Thread: lossless handoff session 1 session 2 session 3 Done · Not done · Next session 3 takes the thread,past ones stay as history Rules apart from the engine Contracts and checksown repo, outlive the engine runs Engine: session orchestratorswappable Log check → snapshot → act → rollback Boundary Second Macfull rights inside your “yes” outside Current Harness roles and flows The human gives tasks to the manager, the manager passes them to the dispatcher, the dispatcher hands work to roles, roles bring in helpers. Cards go from the manager to the human. Thread handoff: session 1, 2, 3. Rules live apart from the engine. The log records every action. The second computer is locked behind your “yes”. People and roles Human Manager Dispatcher Role Role Role helper helper helper card Thread: lossless handoff session 1 session 2 session 3 Done · Not done · Next session 3 takes the thread, past ones stay as history Log check → snapshot → act → rollback Rules apart from the engine Contracts and checksown repo,outlive the engine runs Engine: sessionorchestrator · swappable Boundary Second Macfull rights inside your “yes” outside
It already works
25roles on the team
4games built by agents, open to everyone
7more in private preview

We're opening it up. Leave your email and get it first.

It lives on your machine: plain markdown, one CLI, a board on your phone. For everyone drowning in their agents.

One email when we open. No spam.