Plans: multi-step delegation

When a job is too big for a single chat, askTheodor turns it into a Plan: a set of tasks wired together with dependencies that the app runs for you after you approve it.

What you’ll do

1. Have Theodor draft a Plan from a goal.
2. Read the Plan as a dependency graph.
3. Approve it and watch the scheduler advance it.

1. Let Theodor draft it

You rarely build a Plan by hand. Hand Theodor a meaty request — “scout these URLs, summarise each, draft a post per network, get my approval, then schedule the approved ones” — and he calls propose_plan: a chain of tasks, each assigned to the right worker, wired in the right order.

2. A Plan is a DAG

Under the hood a Plan is a DAG — a directed acyclic graph: tasks point to the tasks they depend on, with no loops. A “write the press release” task can wait on “research the market”; two independent tasks run in parallel. The shape decides what’s allowed to start when.

3. Approve, then it runs itself

A Plan never just starts. It begins in pending_approval and waits for your green light — that’s the safety contract. Once approved, a background scheduler does the work:

• It dispatches each task whose dependencies are all satisfied.
• If a task fails, it retries with exponential backoff up to a budget, then marks it failed.
• When every task is done, it marks the Plan done and writes a final report, with a 📂 Deliverables list linking the documents into the Library.

Plans run even when you’re away

Because the scheduler drives a Plan rather than you sitting in a chat, it keeps advancing in the background — and can run on a 24/7 machine. Budgets and the kill-switch still apply.

Scope the outcome, not the graph

Don’t design the task graph yourself. Describe the result you want — including any “get my approval before X” checkpoints — and let Theodor pick the workers and wire the dependencies. The 📋 Plans screen shows live progress (“4 of 7 tasks done · 57%”).

Three jobs worth turning into a Plan

A Plan earns its overhead when the work has real stages — where step three can’t start until step two is done, and you don’t want to be the one shuttling between them. Three that fit:

A launch you’d otherwise juggle for a week. “We’re launching the team-licensing feature. Research how three competitors describe team pricing, draft the pricing-page copy, draft an announcement post per network, then stop and get my approval before anything is scheduled.” The research finishes before the copy starts because you said so; the two draft tasks run in parallel; and the approval checkpoint is a task in the graph, not something you have to remember.

A monthly report that assembles itself. “Pull last month’s numbers from the spreadsheet, summarise what changed against the previous month, check the Library for anything relevant we shipped, and write a one-page update with a short ‘what needs attention’ section.” Four dependent steps, one report at the end, and the deliverable lands in the Library rather than in a chat you’ll lose.

A backlog you’ve been avoiding. “Go through every page in the docs folder, list the ones that mention prices or features that have changed since v0.9.200, and propose a corrected version of each — one task per page, and don’t edit anything without showing me first.” This is where parallelism pays: independent pages progress at once, and the failure of one doesn’t stall the rest.

Where Plans genuinely disappoint

A Plan doesn’t rescue a vague goal — it just fails in more places. If you can’t describe the finished result in a sentence, the scheduler can’t either. And a long chain compounds small errors: an early task that misreads its brief hands wrong material to everything downstream. Put your approval checkpoints where the cost of being wrong starts to rise, not at the very end.

Recap

A Plan is a DAG of tasks the scheduler advances on its own after you approve it once — dispatching ready steps, retrying failures, and finishing with a report and deliverables. Next: zoom into the single task and the review step that gates “done.”

📖 Reference in the docs: Plans (screen) · Plans (concept)

➡️ Next: Tasks, review & done