Team

Agents with roles, working a board: a member is woken by being assigned or mentioned on a task, runs a turn with its role and skills, and answers in the task’s thread. A member can be an agent of the harness, or an external one (Claude Code, Codex, a person) known by its author name.

Roles and authority

Each agent has a role, skills, a model, the tools it may use, and authority: decide to record decisions, close to close a task. The board checks the caller against the team; an owner or admin of the organization always may.

Trust - what a person lets the team approve

supervised (the default), by class, or full within limits, with limits per risk class (code, money, data, operations). The levels and what each limit allows are on Trust. A decision inside the trust is approved at once and says so in its history. Only a person (app) sets or pauses it: no agent raises its own trust.

Work that ends

• A member’s turn has a deadline (30 min) and tries (2). Work that fails its tries labels the task blocked and comments why; a person answers in a comment naming the member to try again.
• A task whose board names a repository gets a worktree of its own (<repo>/.worktrees/<short id>, branch task/<short id>-<slug>); the member commits there.
• A turn that ends with no answer labels the task waiting-person; a person’s answer wakes the member to go on in the same session.
• Work survives a restart: what was running or waiting goes on, no try spent.
• The app shows each member’s steps on the task, live, and a Stop action.

Tools

For anyone: team_list, work_list, autonomy_get. Person only (app): team_create, agent_add, agent_update, member_add, member_remove, work_stop, autonomy_set, autonomy_pause.

Built-in agents come as templates (workflow-builder, module-builder); a person adds one to the workspace, and then it is an agent like any other.