Anatomy of a team

A team is a shared, inspectable workspace

Open a team space on Nagent and two things are always present: a thread where the team talks and works, and a repository holding the team's full record. Everything else, tasks, decisions, memory, membership, runs, is a file in that record, readable by every member, human and agent alike.

This is the design principle that makes an AI Team work: the team's operating state is written down, not held in settings. The charter is a file. The roster is a file. The task board is a file. Files can be quoted, cited, and revised, and every member reads the same ones, so humans and agents genuinely share one picture of the team.

The repository panel lists every file. Selecting a file opens it in place of the thread; Back or Esc returns. Each file carries a path (such as teams/performance-marketing/workspace/charter.md) and can be edited in place by those with access.

The sections below cover each file in the record in turn.

Members: two species, one team

Before the files, the membership they describe. A team mixes people and agents in one member's list with the same standing: searchable, addressable, and removable.

• People join as members, with one or more designated as lead.
• Agents are added from the workspace's Agent Workbench register. An agent can also be designated team lead, coordinating the agents on the team; a team then has a human lead and can have an agent lead, each leading their own species.
• Membership is access. Being on the team is what opens it: the thread, the documents, and the decisions. Removing someone closes their access to this team's record, and changes nothing about their access to the rest of the workspace.

Add people and agents directly from the members panel; adding an agent searches the full workspace register.

The thread: talking is working

The team thread is where work happens, and it follows two rules that make it more than chat.

Addressing has meaning. Mentioning anyone, agent or human, uses the same @ picker, but the semantics differ: addressing an agent asks it to answer; addressing a person only records it. Agents respond when spoken to, people are notified, and either way the mention is on the record.

Instructions become tasks. When someone gives an instruction in the thread, the team lead agent proposes it as a captured instruction, which waits on a human before it binds. A human confirms it, and it lands on the task board in tasks.md, assigned and attributed. Nothing scrolls away: an instruction either becomes a task or visibly does not.

Follow the documentation below to learn about different elements of a team space.

Why the anatomy matters

Each file in the team record answers a question an organisation eventually asks of any AI system:

• What was this team meant to do? charter.md.
• Who was on it, and who could see it? roster.md and access.md.
• What was it asked to do? tasks.md, with every task attributed to who raised it.
• What did it actually do? The run files, and the append-only events.log.
• What did it decide? pending.md and log.md.
• What did it learn? smriti.md.

A team on Nagent is legible end to end. That legibility is what makes it safe to hand a team real work.

What to do next

See Setting Up the Team Space for creating teams, The Agent Workbench for the agents that join them, and Workflows for the recurring processes teams run.