How the plan is built
A durable server job runs six steps: brief, research, analysis, stages, plan, save. It survives a closed tab, keeps finished work, and retries with a leaner request.
Updated 2026-10-13
When you confirm the dossier, Pythe starts a plan build. It runs on the server, not in your browser, so closing the tab does not stop it.
The six steps
| Step | Label you see | What it does |
|---|---|---|
| 1. brief | Reading your brief | Fixes your stage on the single ladder and stores it on the workspace, where Home, Plan and chat read it. |
| 2. research | Researching your market | Uses the research from setup if it exists. Otherwise runs it. If research fails, the build carries on without market data rather than losing the plan. |
| 3. analysis | Finding your strongest market position | Works out the gap and positioning from your brief and research. |
| 4. stages | Building your next two stages | Writes your current stage and the next one, phase by phase. |
| 5. plan | Drafting your operating plan | Writes milestones across the seven tracks for your stage. |
| 6. save | Saving your company plan | Saves the plan to the workspace. |
The progress bar is weighted by how long each step really takes, roughly research 30, analysis 14, stages 28, plan 24, and small shares for the first and last steps. It is not a fake ticker.
Built to survive a slow model
- A lease. The worker renews a lease every 15 seconds. If it dies, the job is noticed as stalled within about 45 seconds and the step is resumed.
- Finished work is kept. Each step’s output is cached on the job. A retry resumes at the step that failed and does not redo research or spend your quota twice.
- One call per phase. Each roadmap stage phase is written by its own call. If one phase finishes and the other fails, the finished phase is kept.
- Empty is not accepted. A thin or empty phase is refused inside the provider fallback, so the turn passes to the next AI service instead of failing the job.
- Leaner retries. After a failed try, the next one asks for a leaner phase so a slow model still finishes inside its time cap.
- One job per roadmap. The job has an idempotency key, so pressing the button twice, or opening two tabs, returns the same job.
Only a successful build counts as one roadmap generation against your plan. See Credits.
While you wait
The “Building your plan” card slides to one side and a short carousel of tips teaches the roadmap, Pytheus, agents, the canvas, the Library and approvals. The dots, arrows and arrow keys move it by hand.
If it fails
- You see a plain message and a Try again button. Your answers are still saved.
- A failed step shows which one failed, and what finished before it.
- You can also follow a build from your own agent with
get_plan_build. Starting a build is always done by you in Pythe. See MCP: plan tools.
Related
Previous: Market and competitor research · Next: Workspaces, members and invites