Plans, subtasks and delegation

When a request is bigger than a single reply, the assistant stops improvising and organises. You can watch the whole thing happen.

Where it runs: in a task, and in AI staff runs.

What the assistant can do

It can	What that looks like
Write a plan	A todo list in the task’s Plan section, ticking off as it works
Spawn a subtask	Hands a self-contained step to another AI staff member and waits for the result
Delegate a task	Creates a separate follow-up task for another staff member — fire and forget
Create a task	A new work item in the workspace for someone to pick up
List tasks	Checks your open tasks, mostly to avoid creating a duplicate
Post into the thread	Shares progress or an intermediate result with the humans watching
Schedule itself	Comes back later with a given instruction — a “check back” trigger
Wait	Pauses briefly for something just triggered to settle

The plan and the delegated subtask are both visible on the reply; the tick means the subtask reported back.

The plan

For anything multi-step, a todo list appears in the task’s Plan section before the work starts. It is not decoration — it is what the assistant is actually working from, and items tick off as they complete. If the plan is wrong, say so immediately; correcting a plan is far cheaper than correcting the work.

Subtask or delegation

The difference is whether the result is needed now.

A subtask is a sub-agent doing a self-contained piece of the current job — research, a draft, a check — while you wait. It gets its own thread, does the work there, and reports back. You see one coherent reply in your conversation instead of a wall of sub-agent chatter, and you can open the subtask from the Plan section to read the working.

A delegated task stands on its own after this conversation. It shows up in the task list assigned to another staff member, runs by itself, and returns nothing to the conversation that created it.

Scheduling and waiting

A staff member can defer its own work — “check whether the supplier replied tomorrow morning” — by scheduling a later run with an instruction to itself. Short pauses (a few seconds, waiting for something it just triggered to settle) are a separate, much smaller thing; anything longer becomes a scheduled run rather than a held-open session.

Working with it well

Correct the plan, not the output. The plan is visible before most of the cost is incurred.

Say when something is a follow-up. “Do the analysis now, and have someone draft the client note afterwards” maps cleanly onto subtask plus delegation.

Expect a report, not a transcript. Sub-agents summarise. If you want to see how a conclusion was reached, open the subtask.