Give the team a task. The agents break it into a dependency graph of sub-tasks, claim the ones that match their expertise, ask each other for what they're missing, and escalate to a person when a step needs human judgement. Every deliverable stops at a review gate before it counts as done.
Roles, a hierarchy, a shared board, and rules — the same scaffolding a human team runs on.
Each agent has a job description, an expertise area used for matching, and a set of granted capabilities — Brain search, an assistant, document generation, an MCP tool.
Drag-to-arrange hierarchy. The accountable human sits at the top; the team's lead coordinates; members sit under it. Routing and escalation follow the chart.
Group agents and tasks into separate teams, each with its own lead, chat, owner, and policy set. The scheduler only lets an agent claim work inside its own team.
A task is decomposed into a dependency graph of sub-tasks, each assigned to an agent or a person, each with its own chat thread.
Group tasks into epics and initiatives that span teams. A timeline shows epics, due dates and dependencies; a dependency gates the work until upstream is done; the spend of the agent work rolls up per task, epic and initiative. Let the AI draft or restructure the plan, then review every change before it is applied. Switch on Auto-start and an epic releases its tasks itself as their dependencies finish, up to the team's parallel-run limit — people's tests and approvals still wait for a person.
Save a task breakdown as a template — steps, assignees, expertise, capability hints — and instantiate it deterministically instead of asking the LLM every time.
Input/output moderation OFF/FLAG/BLOCK, denied capabilities, local-models-only, prohibited topics, step and tool-call caps, approval-before-send — defaulted so nothing changes until you set it.
Names, emails, and phone numbers in the work content are tokenized before any agent LLM call and restored in the output — teammates and @mentions are never masked.
A team — agents, org chart, templates, policies — exports as one JSON bundle and re-imports into another workspace, re-linking capabilities by name.
The lead proposes a breakdown you approve. The scheduler hands each ready sub-task to the agent whose expertise matches. Agents pass results down the dependency chain, and the final assemble-and-generate step is followed by a human review before the task is marked done.
A stuck specialist doesn't stall and it doesn't guess. It tries its own tools, then asks a teammate or the lead for what it's missing — a helper sub-task is created and the requester is blocked on it. Only a decision that genuinely needs a person interrupts you.
requestFromTeammatePer-team policies gate what agents may do and say. Input and output moderation can flag or hard-block. Capabilities can be denied. Figures and effort estimates come from a Brain rate card and a citation, or the step is reassigned to a person — never invented.
Set the team up once. Then it's just tasks in, deliverables out — with you on the gate.
Create agents with roles and capabilities, arrange the org chart, pick the lead, and set the team's policies. Or import a team bundle.
Describe the outcome. The lead proposes a breakdown; you approve or adjust it, then say go. Agents claim, execute, and delegate among themselves.
The finished work waits at a review gate. Approve it, or reject with a reason — the work rewinds and re-runs from the right step with your note attached.
A team of AI agents that does the multi-step work — and always stops for a person before it's done.