Before the agent writes a task

An assistant that turns your paragraph straight into forty tasks is guessing, and the guess is expensive because it looks like work. Foreweave's agent is deliberately slower at the start: it will not create tasks ad hoc until you have agreed how it intends to plan. Everything below is the road to that agreement.

1. The project profile

A profile records what kind of programme this is: the domain, the delivery model, which approvals or certifications gate it, and how much of it is genuinely novel. From that, Foreweave derives a complexity rating — and always shows it alongside the factors that produced it, so you can disagree with the reasoning rather than the number. Every approval you record maps to an actual gate in the eventual schedule; approvals that map to nothing are the ones that surprise you in month nine.

Your organization can publish profile templates so a new project starts from your standard shape instead of a blank form, and a line of business can carry its own.

2. The kickoff interview

Where the profile has gaps, the agent asks. It works from a defined minimum dataset — the set of facts below which planning is guesswork — and it will not begin proposing a guideline until that set is complete. Three things make this bearable:

• It never re-asks something you already supplied, in the form or in an earlier session.
• You can stop halfway and come back; the interview resumes where it was.
• Some answers are legitimately "none" — no external vendors, no compliance regime — and saying so explicitly counts as answered. Silence does not.

Every captured field keeps its provenance: whether you typed it, it came from an uploaded file, or the agent inferred it. What stays genuinely unknown is written into an open-questions note rather than quietly assumed, so the gaps travel with the plan.

3. The planning guideline

Now the agent proposes how it will plan: the phases, the gates, where the approvals land, the shape of the breakdown. You accept it, amend it, or reject it and ask for another. Until a guideline is accepted, the agent cannot create tasks ad hoc.

Two deliberate exceptions: importing a file and building the plan directly from an accepted guideline both bypass the gate, because in each case you have already told it the structure. What the gate blocks is invention.

4. Memory

Decisions, constraints and lessons persist per project, so the next conversation does not start from nothing — the agent reads them before it reads your message. Two boundaries are absolute:

• Memory never crosses a project or an organization. There is no shared pool.
• Promoting a lesson to your wider line of business — "on programmes like this, certification always takes longer than the vendor quotes" — takes an explicit human acceptance, and identifying detail is stripped on the way.

You can read everything the agent has remembered, and delete any of it.

Then, finally, the plan

With a profile, a complete dataset and an accepted guideline, ask it to build. The draft arrives as a scenario overlay you review as a diff — see drafting with the agent for how to read it and what it will refuse to invent.