Private — dogfood workspace
3 humans · dozens of agents
AI operating system — one-person and small-team companies
Xiaoyin Qu
Timeline
2026-06-11Leverage tracking goes live for the Tycoon workspace. First snapshot: 46× (1,000 human-equivalent hours delivered vs 21.5 chairman hours in a 7-day window, 400 tasks done).2026-06-19June peak: 522×. 7,206 hours delivered vs 13.8 chairman hours. 1,201 tasks completed in the window.2026-07-07July ramp crosses 266×: 4,086 hours delivered vs 15.3 chairman hours, 681 tasks.2026-07-14Highest-volume week: 716×. 10,170 human-equivalent hours delivered — roughly what a 50-person company ships in a week — against 14.2 hours of founder time. 1,018 tasks.2026-07-171,261 tasks completed in the 7-day window — the busiest task week on record.2026-07-21Peak leverage: 917×. 5,320 hours delivered vs 5.8 chairman hours, 532 tasks. The founder's week cost less than a lunch break per day.2026-07-23The team's real operating rhythm shows: a 244× → 24× dip over four days as work paused, then recovery.2026-07-31Month end: 72× with 2,670 hours delivered vs 37.4 chairman hours — sustained, not a one-day stunt.Key insights
- 01Leverage is a people-math story, not a speed story. The Tycoon ROI Equation prices what a real org would burn (execution hours × coordination tax × interruption tax) and divides by what the founder actually spent. The machine cost is a rounding error — the leverage comes from collapsing two human multipliers: coordination (n(n−1)/2 sync channels) and interruptions (every ping that stops the work).
- 02917× is the July peak: 5,320 human-equivalent hours delivered against 5.8 chairman hours in the 7-day window ending July 21. The 3-person team (Xiaoyin, Richie, Darren) kept their roles — one owner per decision, one approval channel — while the work ran through Tycoon Agent.
- 03The highest-volume week was July 14: 10,170 hours delivered at 716×. That is roughly a 50-person company's weekly output, shipped by three people and their agents.
- 04Scale is repeatable, not a screenshot. The workspace ran 400+ tasks a week in June and 1,000+ in July; the busiest week completed 1,261 tasks. The dip-and-recover rhythm (July 23–26) is what real operations look like — no demo cleanup, no cherry-picked week.
- 05The ramp is the product working. 46× (June 11) → 522× (June 19) → 917× (July 21) tracks the team teaching Tycoon Agent how the company works: who owns what, what good looks like, what needs approval. Every corrected task made the next one cheaper.
- 06Git tells the same story. The tyny repository alone recorded 5,916 commits in July 2026, the majority authored by agents working under the chairman's identity — work that would have been a multi-team effort in a conventional company.
Stack used
Tycoon Agent — the AI manager: task orchestration, agent routing, approval flow, verification and reportingAgent roster — coding, marketing, QA, data and design agents plus bring-your-own Codex / Claude Code / Hermes agents1,000+ connectors — GitHub, Stripe, Slack, Notion and the rest of the company's tool surfaceCompany Brain — docs, company skills and memory that every agent readsHuman layer — 3 people: product (Xiaoyin), engineering (Darren), production access and ops approvals (Richie)
What this means for you
- One manager, not one agent per task. The team directs Tycoon Agent, not individual agents — one conversation replaces a coordination graph.
- Human attention is the only scarce resource. The metric that matters is chairman hours, not tokens spent. Tycoon Agent's job is to spend your attention only at real decision points.
- The first month is a teaching month. Expect the ramp (46× → 500×+), not the peak, on day one: every correction you make teaches the manager how your company works.
- Verify the work, then trust the output. The team's rule: work ships only with proof — changed files, tests, live URLs — reviewed against the brief.
- The numbers are auditable. The leverage figure is a documented equation over production data (7-day rolling window, task-level human-equivalent hours), not a marketing dashboard. Read the methodology before quoting it.