Drafting with the agent

Describing a program

Give the agent what you'd give a new planner on day one: the phases, the order they happen in, the hard dates, and the constraints that aren't negotiable. For example: "Three build sprints of two weeks, integration testing after each, certification after the last test pass, launch can't slip past Oct 24." Structure beats length — phases, sequence, dates.

What it drafts

The agent creates tasks, links, milestones, and estimates through the same validated engine API you use — its edits obey calendars, reject cycles, and report conflicts exactly as yours do. Everything lands in a scenario as a reviewable diff: task by task, link by link. Your baseline is untouched until you promote.

What it refuses to invent

When it doesn't know a duration, it asks instead of fabricating one. Quantities it can't derive from your description — team size, a vendor's lead time, the certification queue — come back as questions, not guesses. A drafted plan with honest gaps beats a complete plan built on invented numbers.

Simulations you can cite

When the agent runs a Monte Carlo for you, it pins the seed and reports it, so any number it quotes ("certification is critical in 84% of runs") can be reproduced exactly from the run record.

Trust boundaries

• Off by default for every tenant; an admin opts in, with disclosure that an external model provider is involved.
• Zero-retention model endpoint; your data is never used for training.
• Configured confidential identifiers are redacted before transmission and restored on return.
• Every prompt, response, and tool call is audited per tenant, with token and cost usage reported per user and per organization.

The full data-handling commitments are on the security page.