Plan with the Agent

Building a plan by hand means deciding the outcomes, the work under each one, the tasks under that, and the dates — before you have written anything down. The Agent can draft all of it from a goal, grounded in the research already in the project, and hand you something to edit rather than a blank outline.

Nothing it drafts is saved until you say so.

Ask for a plan

Open the Agent on the Plan tab — the composer asks What are you trying to accomplish? — and describe the goal in your own words. Or type / anywhere in the Agent and choose Break down a goal.

Either way you get the same behaviour, because the Agent reads the plan you already have and works in two stages.

Stage one: agree the outcomes

The Agent replies with a single table of three to six milestones:

Each one has to end in something demonstrably true, one accountable owner, and a date range. Then it stops and asks you to confirm or adjust — which to drop, split, or re-order, and whose name belongs where. It deliberately does not expand into activities yet, because reworking the outcomes is cheap at this point and expensive two levels down.

Stage two: expand into work

Once you confirm, the Agent repeats the milestones table and adds a second one underneath:

Every activity names the milestone it belongs to, carries a kind (meeting, workshop, deliverable, audit, or other), and lists the to-dos inside it. Keep refining in the conversation — the Agent re-emits both tables in full each time, so what you save is always the latest version.

Steps that leave something behind

The Agent treats each activity as a job and each to-do as a step, so a step that should produce a durable artefact names it in square brackets — Map the journey [Current-state journey (journey)]. Those become deliverables on the saved activity, ready to fill by linking an existing artefact or creating a new one. Not every step gets one, and you can add, rename, or waive them afterwards.

Review before it lands

When the reply contains a plan, Save & submit plan appears under it. Clicking it does not write anything: it opens a preview headed Review the plan before saving, showing the milestones with their activities and to-dos nested underneath.

In the preview you can:

• Uncheck any milestone, activity, or to-do you do not want. Unchecking a parent takes its children with it.
• Edit titles, dates, and owners inline.
• Read what each activity delivers — the preview lists the deliverables under the activity and the step they belong to. They are read-only here; edit them on the activity once it is saved.
• Watch the count at the bottom — milestones, activities, and to-dos — before committing.

Save & submit plan in the preview writes it. Items with the same name are updated rather than duplicated, so refining an agreed plan and saving again adjusts what is already there instead of creating a second copy of it. Once it lands, the button reads Plan saved.

Just the next steps

Sometimes you want tasks, not a plan. The Generate to-dos slash action asks for concrete to-dos under a Next steps heading, and the reply carries a Create to-dos button with its own preview. Those land in the project To Dos list, and you can attach them to activities later.

Re-planning

The Agent reads the current plan before answering, including what is scheduled and what is still in the unscheduled tray. That makes re-planning a normal conversation: tell it a date has slipped, or that a milestone is no longer needed, and ask it to re-sequence. Because saving matches on title, the result updates the plan you already have.

From a connected assistant

Claude, ChatGPT, Grok, and Build IDEs connected over MCP can do the same thing from their own chat window using list_plan, upsert_activity, and create_todo. They are instructed to propose the plan in the chat first and write only once you have agreed it.

Next

• Plan people & owners
• Applying and materialising answers
• Reviewing AI output