How Pythe works
One chain connects it all: onboarding, company facts, research, roadmap, Pytheus, employees, tasks and files. Here is each link and where it lives.
Updated 2026-10-13
Every part of Pythe reads from the same source: the facts you gave it about your company. That is why a researcher, a writer and an engineer all start with the same picture of you.
The chain
| Link | What happens | Where you see it |
|---|---|---|
| Onboarding | You describe the company. Pythe drafts the fields and you correct them. | Onboarding |
| Company facts | The confirmed dossier is saved on the workspace. Every agent prompt starts from it. | Workspaces |
| Research | Pythe researches market, audience, competitors and risks, with sources. | Market research |
| Roadmap | A stage-aware operating plan: tracks, milestones, steps and graduation criteria. | The plan |
| Pytheus | Your AI cofounder. It reads the plan and the Library, answers, runs code and briefs the team. | Pytheus |
| Employees | Specialists with their own computers and tools, each following a method for the job. | The AI team |
| Tasks and runs | Work is tracked: a brief, a definition of done, progress, a check, and a report. | Tracked tasks |
| Files and approvals | Results land in the Library. Anything that leaves waits for you. | Approvals |
What runs where
- The app is a React single-page app. It reads and writes your workspace, and streams what the team is doing.
- The backend is a set of Supabase Edge Functions plus a Postgres database with row-level security. Long jobs, such as the plan build and agent runs, are durable: they run on the server, hold a lease, and are resumed if a worker goes quiet.
- AI models are called through a provider fallback. If one service returns a well-formed but empty answer, the turn passes to the next instead of failing.
- Agent computers are cloud sandboxes, one per agent, that start when work needs them and sleep when idle. See Agent computers.
- Connectors such as Gmail and Slack run server-side with your own tokens, never in the browser.
One piece of work, end to end
- For example, you ask in HQ: “Find 15 indie cafe owners in London and draft intros.”
- Pytheus decides the Outreach employee should do it and starts a run with your brief.
- The employee reads your company profile and shared memory, writes a short plan with success criteria, then does the work on its computer: searches, opens pages, builds a CSV of leads with a signal and a source URL for each.
- It drafts the emails. Sending one through Gmail creates an approval card with the exact text. Nothing is sent yet.
- Before reporting, it checks its own work once against its criteria, fixes gaps, then writes a report that separates what was sent from what is only waiting for you.
- You approve, edit or reject each email. Approved ones go out from your Gmail with an opt-out line.
- The CSV and the drafts are saved to your Library, and the run is added to the team’s shared memory.