Agent Studio : AI agents that work as staff - and wait for people.
Build agents that work across the platform: the studios are their tools, and they are the studios' tools too. Every agent acts as someone, every tool says what it does to the world, and anything that changes something waits for a person.
What it does
Agent Studio, point by point.
It acts as someone
An agent works with the access of the person it works for, or with an account of its own. It can never see more than that.
Tools declare their effect
Each tool is marked: it finds, it proposes, it makes, or it changes. What changes something is held until a person approves.
A member of staff
An agent can have its own account, a supervisor, a duty that can be paused, budgets, routines and a timesheet of the time it worked.
Approvals in the cartable
A held action reaches the supervisor's cartable with what the agent wants to do and why. Approve, decline, or answer.
Reachable from everywhere
In a chat, as a step of a process or a flow, from an app, from the assistant - and from outside through an MCP endpoint.
Runs that survive
A run that waits for a person or a reply is not a held thread: it ends, and takes up again when the answer comes - through restarts too.
One switch
Every agent on staff can be paused at once.
See it work
Be the supervisor
An agent runs until it would change something. Then it is your decision.
IT helper
An agent on staff. Its supervisor: you.- findsReads the request: "I cannot connect to the VPN."
- findsSearches the help articles the person who asked may read.
- proposesDrafts an answer from the article, with its three steps.
- changes Wants to reset the VPN account of the person who asked.
This changes something, so the agent stops here and asks its supervisor.
- done
Getting started
Three steps in.
- Open Agent Studio and describe the agent: what it is for, and for whom.
- Check its tools. Each one shows its effect.
- Try it in the studio, then publish it and choose where it may be reached.
Questions
What people ask about Agent Studio.
Can an agent act without anyone knowing?
No. Every run is recorded with its steps, tools and cost, attributed to the agent and to whoever it acted for.
Which models does it use?
The ones you configure: Anthropic's models, any OpenAI-compatible endpoint - one you host included - and the Claude Code, Codex and Gemini command-line engines on your own runner.
Can an agent use a password?
It can use a credential from the Vault for the work it was granted - and never reads its value.
Works with
It does not work alone.
One assistant beside every page - that proposes, and obeys.
Everything waiting for you, in one list.
Credentials that are used, never shown.
Connect the systems of the organisation.
Courses written, examined and tutored by a team of agents.