Skip to content
MenuClose
Register interest

A desktop workspace for Claude and Codex.
Project context, agent sessions and human review.

Skip to documentation
Agent teams

AigentFlow Desktop is in pre-release. This documentation reflects the current build and will change before general availability.

Working with agents

Agent teams

One lead agent talks to you. It delegates to specialists, has their work checked by a different specialist, and reports back.


Teams are off until you turn them on. When they are on you stop talking to whoever happens to be in the tab and start talking to a lead. The lead plans, delegates, monitors, verifies and reports. It does not implement the work itself.

The roster

Twenty professions are available across six groups: Leadership, Product and Design, Engineering, Data, Quality and Delivery. Four are enabled out of the box.

MemberWhat it does
Tech LeadPlans, delegates, verifies, reports
Frontend EngineerReact, CSS, accessibility, component work
Backend EngineerServices, IPC, main-process logic
Code ReviewerRead-only review, APPROVE: YES or NO with evidence

Each member has its own provider, model and effort level, so a reviewer can run on a different model from the engineer whose work it is reading. One Leadership profession must be enabled to act as the lead, and member names must be unique.

How work is handed out

The lead proposes a plan as a set of tasks, each naming one member. There are two shapes.

  • Deliver changes code.
  • Investigate is read-only and comes back as a report.

A member works on one task at a time. If a member is busy when the plan starts, its task is skipped and the lead is told. After a deliver task completes the lead has it verified by a Quality member, and it may never nominate the member that did the work.

Approval

By default you approve the plan in the app before any member starts. A separate setting, Autonomous goals, lets the lead start its own plan as soon as it has proposed one. Every auto-approval is logged. That setting is global, and a per-project team override cannot turn it on.

What a member is given

A member does not inherit the conversation. It receives the approved plan, the project instructions from CLAUDE.md and AGENTS.md, and the project context files, and nothing from the discussion that produced its task. Its instructions are deliberately narrow.

  • Work through the steps in order.
  • Do not replan, restructure or propose an alternative approach.
  • Do not add features, refactor or clean up beyond the plan.
  • If a step fails, stop and report it rather than working around it.

If a member starts arguing with its own plan, the app recognises the phrasing and halts it, marking the outcome as a deviation. That is a containment, not a failure.

A member can never outrank the session

Members are capped at Working. A lead in Discuss, Plan, Review or Guarded starts members in that same mode. A lead in Working or Autopilot starts members in Working, so the unattended powers you consented to stay in the tab you consented in.

Watching the work

Each member the lead puts to work gets a card showing its state, how full its context is, its profession, model and effort, and the task it is on. Opening a card shows that member’s session read-only: you watch, you do not type. You talk to the lead.

When a member finishes you can open its execution report: the outcome, how long it took, each step, and the files it modified.

Delivery

One setting decides what happens to finished work. Ask you before any merge, open a pull request for you to review, or leave it on a local branch.