Agents at Work on Tasksnew · 1.6

The Agents tool lists the user's agents in groups. An agent is a specialist with a name, an emoji or icon, a color, a model, an autonomy (autonomous, or asking before what acts) and its own agent.md: who it is, what it knows and how it works. A group can carry a context every agent in it shares. A click on an agent opens its definition in a tab; its form also sets what it may do (Access) as a supervisor and as a specialist.

A task of a project can be given to an agent, from the task form, the task menu or the backlog. Set to Working, the agent starts: Claude Code runs headless in the repository of the project, with the agent's definition, the project context and the task (its context, assets and latest comments) as its opening message. One run works per repository at a time; the rest wait in a queue, under a global limit of runs at once (Settings, Agents).

The Activity tab of the task shows the run as it is: its state, agent, model, what it did last and for how long; its plan (the steps it set itself, with those done ticked); a timeline of every event (what it read, ran, wrote, asked); earlier runs folded under it; and a box to tell it something while it works. What it asks shows as cards answered right there: a permission (Allow or Deny), a question, or a run that stopped and can be retried.

A run sets its task to Review with a comment of what it did, and is done. It can ask the user with anxeb_task_ask (the task shows Waiting, purple, until the answer comes as a comment, which resumes it). A run never closes a task: the user, or the supervisor, does. Stop and Reset are in the task menu; giving the task to another agent asks first, and a task taken back to Pending stops its run.

Every run is told in Notifications: "{agent} needs you" for a permission, a failure or a run that stopped without Review, with the sound of asking; "{agent} finished {task}" when it reaches Review (Settings can turn this off); "{task} waits for your answer" for a question. The backlog shows the light of each run before its title: green beating while it works, grey while queued, red while it needs the user.

What it does

• Agents in groups, each with its agent.md, model, autonomy, emoji or icon and color; group contexts.
• Headless Claude Code runs per task, in the repository of the project, through a queue (one per repository, a global limit).
• Activity: state, plan, timeline, earlier runs, a box to talk to the run.
• Permission, question and failure cards answered in place; Always ask rules in Settings.
• Waiting state while a task waits for the user's answer.
• What an agent may do, layered: its own access, the project's, the task's.
• Notifications with sounds for what needs the user and what is done.

How it behaves

• Runs survive a restart: one that was working resumes its session.
• A run cannot close its task, give it to someone else or change its code.

Why it matters

• Hand a ticket to an AI specialist and review what it did, without leaving your terminal.
• Every step of the agent in view: nothing happens behind your back.