Review the plan

The plan is the cheapest moment to course-correct. Read it before the Composer burns credits.

The 30-second version

When the workroom is in planning, the Planner is still working. As soon as it produces a plan, the workroom transitions to executing automatically. To intervene, Pause the workroom or Replan afterward. Each task in the plan shows its title, description, assigned agent, dependencies, and rough time estimate.

Why this matters

Plans are easy to read and cheap to fix. A bad plan that runs to completion costs you the full workroom credit — usually tens of credits. Catching a missing section or a wrong agent assignment at the plan stage costs nothing.

There's no separate "Approve plan" button today: the workroom auto-executes once the Planner finishes. The path to intervention is Pause while planning is still in flight, or Replan to discard non-completed tasks and rebuild.

How it works

Reading the plan

On the workroom detail page, the Tasks panel shows the full task list once the Planner finishes:

• Ordinal — execution order.
• Title and description — what the task does.
• Type — agent_work, composition, approval, tool_call, wait, user_input.
• Assigned agent — system Planner / Composer or a custom agent.
• Depends on — which earlier ordinals must finish first.
• Estimated duration — rough seconds, advisory only.

The last task is always a composition task — that's the one that produces the deliverable.

[SCREENSHOT: workroom task list showing planner output]

What to look for

A plan is on track when:

• The deliverable is broken into sections that match the brief (Exec summary, Scope, Pricing, Timeline — for a proposal).
• Each section is assigned to a sensible agent (a B2B-tuned custom agent for the SaaS proposal, the default Composer for a generic recap).
• Dependencies make sense (Pricing depends on Scope; Timeline depends on Scope).
• The final composition task pulls everything together.

A plan needs intervention when:

• A required section is missing.
• The wrong agent is assigned (e.g. Composer is doing planning).
• A dependency is broken (composition runs before its inputs).

Three ways to intervene

1. Let it run. Most plans are fine. The Composer will assemble the final output and you'll review the deliverable, not the plan.

2. Pause. From the workroom header, click Pause. The workroom moves to paused. In-flight tasks finish; new ones don't start. Resume by transitioning back to executing.

3. Replan. From the workroom menu, click Replan. AtelyaOS calls the Planner again with the original inputs plus the current state. Tasks marked done are kept; everything else is discarded and regenerated.

When the plan is wrong, why

The most common reasons a plan is wrong:

• Goal was vague. Edit the goal and replan.
• Missing input. A section is missing because the Planner didn't have the source. Add a paste block (you can edit pasted inputs) and replan.
• Wrong agent in the room. If the Planner picked the generic Composer when you wanted a client-specific agent, add that agent to the workroom and replan.

Common pitfalls

• Watching the plan instead of intervening. If the assigned agents are wrong, pause now. Watching three tasks fail in sequence wastes credits.
• Replanning to fix small wording. Replan throws away in-flight Composer work. For wording-level changes, wait for the deliverable and use Request revision instead.
• Editing the goal mid-execution. The Planner already ran on the old goal. To regenerate against the new goal you need to Replan, otherwise tasks finish using the original framing.
• Forgetting that dependencies matter. A plan may look fine in the list view but have a broken dependency graph. Read the Depends on column.

What's next

• Approve agent work — once execution finishes.
• Briefs and plans — Planner schema in detail.
• Workrooms — full lifecycle reference.