Auto-pickup and dependencies

You can drag tasks onto agents by hand. Or you can turn on auto-pickup and let agents claim work themselves, bounded by a sensible concurrency cap and a dependency-aware scheduler.

The model

Three things compose:

1. A concurrency cap. Each agent runs at most two tasks at a time in v1. We’ve kept this fixed deliberately while we learn what a useful default looks like across real teams; tunability comes once we have the data to back it.
2. An auto-pickup opt-in per agent. When it’s on, the agent claims the next ready task on its own whenever a slot opens up.
3. Dependencies between tasks. A task is “ready” when it’s in To Do, has an agent in scope, and every task that blocks it has finished.

The combination: an agent runs up to a couple of tasks at a time, automatically claims the next ready one when a slot opens, and never picks up something whose blockers are still in flight.

Why this matters

Coding work isn’t a flat queue. “Add the checkout flow” depends on “wire the webhook handler”. The dashboard refactor blocks three follow-up bug fixes. Agents that try to work through a queue without dependency awareness will either sit idle when there’s actually unblocked work available, or start work that immediately fails because its prerequisite isn’t done yet.

Vibetight’s scheduler always picks the highest-priority task whose blockers are done. If nothing’s ready, the agent waits, instead of picking up something that’ll just fail.

What you see on the Gantt

The Dependency view (the third tab in the board switcher) shows everything visually:

• Columns are dependency depth. Tasks with no blockers sit in the leftmost column. Tasks blocked by one layer sit one column to the right. And so on.
• Bar color = status. Blue = todo, emerald = in progress, green = done.
• Dashed lines connect blockers to dependents. Following the line shows the path that has to clear before something can start.
• Green #N pill on a todo bar = “this is the agent’s next pickup, ranked #N in its queue.” Multiple agents = multiple #1s, one per agent.

The view is interactive: click a bar to open the task, drag to reorder priorities within the same depth.

Pairing this with Maestro MCP

Plan via Maestro → tasks land on the board with dependency edges → auto-pickup agents start executing the leaves → as they finish, deeper tasks unblock and get picked up.

The whole pipeline runs without you touching the UI. You drop in to review diffs, answer Action Required prompts, and steer when the plan needs adjusting.