Agents

Ask your assistant to dig into a competitor's pricing while you carry on drafting, and the research goes somewhere else: to an agent that does that kind of work, keeps its own notes on it, and reports back into your chat when it is done.

That is what an agent is in Player One. A named member of staff with one role, its own files, its own memory and its own settings. You talk to one the way you would message a colleague, you add more as the work grows, and you call them whatever you like.

The team you start with

A new installation arrives with three agents you will actually talk to.

Your main assistant is the front door. It knows you, answers most things itself, hands work to whoever is better placed, and it is the only agent that can create or change another agent.

A general worker takes routine side work off the main conversation: a quick lookup, a small edit, a piece of throwaway research.

A coordinator plans. When a piece of work is too big for one agent, it breaks the work into tasks, picks an agent for each from what that agent is good at, and drives the job to the end. It never does the work itself.

Behind those sit a few system agents you rarely address directly, including the one that tidies memory while your installation is quiet.

Specialists are added rather than shipped. A catalogue of ready-made roles comes with your installation, covering research, writing, design, engineering, finance, marketing, support and more. Pick one, or describe the job in your own words and let Player One build it: Grow your staff walks through both.

Some agents face your customers instead of you. A sandboxed agent stands alone: it knows only what you gave it, acts only through the narrow tools you built for it, and never joins a team or a job. Vault and sandboxing covers that boundary.

What every agent has of its own

• A role and a set of strengths. The strengths are what a coordinator reads when it picks someone for a task, so plain words pay off.
• A workspace. Its own folder of files, created with it. See Workspaces and files.
• A Memory Space. Its own durable memory, created with it and following it between projects, which is why a specialist gets better at its one job. See Memory.
• A bounded set of tools. Every agent is allowed a deliberate list and nothing beyond it.
• A permission mode. Conservative, Most things or Everything: how much it may run on your installation before it has to ask you.
• Its own voice and model. The voice it speaks with and the model it thinks on are set per agent, not once for the whole site.

The agent's own page

Open an agent and you get its page, with tabs across the top.

Info holds the written instructions that make it who it is. Skills covers what it knows how to do, and carries the strengths editor. Workspace browses its files. Task shows what it is working on. Projects decides which projects it belongs to. Advanced opens on Permissions, what the agent may run on your installation without asking you (see Stay in control), followed by Models, Voice, Tools, Prompt, Memory and Dangerous.

The Tools list under Advanced flags a name that no longer means anything, a tool renamed or withdrawn in a later version, and offers to take it out. It can be removed from that one agent, or from every agent that carries it in a single action, which is the quicker fix when an old name has been sitting in several policies for months.

Instructions for every agent

Rules that apply to all of them live in one place: All agents, in the ... menu on the Agents page. Whatever you write there is loaded by every agent on your installation, ahead of the project's instructions and the agent's own, so keep it short. Every line is paid for on every turn.

Where to find them

Agents sits in the Manage group of the sidebar (/agents). Cards sort by recent use, filter by tag, and search from the box in the top bar. Asking your assistant for a new agent is the usual way; New agent opens the same builder beside the chat, and needs a real project selected rather than All projects.

To see who is working this second, open Active Agents under Manage (/manage/active-agents/by-project). It groups live conversations by project, links into each one, and lets you stop a run without losing the chat.

Teams are set up in conversation rather than on a screen. Tell your assistant who works for whom, "Nadia runs outreach, Bob does the books, both report to my assistant", and it writes a team page every internal agent reads: what the team does, who leads it and a roster. As the staff grows, a manager gets a sub-team page of its own, and a team learns instructions the way a skill does, "always hand invoices to Bob". Customer-facing sandboxed agents stay outside the team by design.

For when to hand work to another agent, and what that looks like in the chat, read Agents, delegation and team chat.

Adding to the team

New agents come from the Forge: a catalogue of 77 ready-made specialists that opens beside the chat, or a role written from scratch when nothing fits. You describe the job, your main assistant settles the shape in one round of questions and builds the agent in one pass. Grow your staff covers when to add one and when not to.