The Maestro, Season 2 ยท Part 7 ยท Phase 6 ยท ๐ฏ CEO hat
Hunting the Next Moat
Today's moat, the customer skill library, will not last forever. Gopala lines up five candidates for the next one and refuses to bet on a single forecast. Instead he scores each candidate in three possible futures, picks the one that does least badly when he is wrong, and checks how long a much bigger rival would need to copy it.
๐ผ CEO๐ฌ Research director๐งญ Strategist
โ ๏ธ How to read this (as of 3 Nov 2026)
moat_futures.py and copy_time.py were run (real output below). Futures, probabilities, scores and copy times are the story's scenario estimates. The method (minimum regret) is standard decision analysis. Disclosure: Claude, made by Anthropic, helps write this site.
Mission log: months 30โ32
The map: five moats, three futures
Choose for the worst case, not the best guess
How long would a rival need?
Sandbox: bet on a future
Five mistakes
Delegation box
Budget and risk tracker
The series
Reading the numbers
1. Mission log: months 30โ32
Week 120 ยท Five candidates
- From the research tracks and the startup notes: environments, skill library, harness design, tool use, long-horizon agents.
Week 122 ยท Three futures
- The board agrees three scenarios and rough probabilities.
Week 124 ยท The pick
- Environments: lowest regret, 18 months to copy.
2. The map: five moats, three futures
๐ธ๏ธ Which moat survives which future
Futures in red, candidate moats in purple, tests in blue. Edge labels are scores (0โ10) or months to copy.
- Three things to notice.
- Only one candidate scores 7 or more in every future.
- The most exciting candidate (long-horizon agents) scores lowest.
- The two durable moats are the ones built on data nobody can buy.
3. Choose for the worst case, not the best guess
moat_futures.py ยท Python 3, no dependencies; scenario scores
"""moat_futures.py โ score five candidate moats under three market futures,
then pick the one that holds up best (minimum regret), not the one that wins
the single most likely future.
Scores (0-10) and future probabilities are the story's SCENARIO estimates.
Futures: A = frontier labs ship strong coding agents fast;
B = open models catch up and prices collapse;
C = regulation makes change-by-change evidence mandatory.
"""
MOATS = {
"verified environments at scale": (8, 7, 8),
"customer skill library (today)": (6, 5, 8),
"automated harness design": (5, 7, 4),
"self-improving tool use": (7, 4, 5),
"long-horizon agents": (4, 3, 5),
}
P = (0.45, 0.30, 0.25)
def main():
best = [max(v[i] for v in MOATS.values()) for i in range(3)]
print(f"{'moat':32} {'A':>3} {'B':>3} {'C':>3} {'expected':>9} {'worst regret':>13}")
rows = []
for m, v in MOATS.items():
ev = sum(p * s for p, s in zip(P, v))
regret = max(b - s for b, s in zip(best, v))
rows.append((regret, -ev, m, v, ev))
for regret, _, m, v, ev in sorted(rows):
print(f"{m:32} {v[0]:3} {v[1]:3} {v[2]:3} {ev:9.2f} {regret:13}")
pick = sorted(rows)[0]
print(f"robust pick: {pick[2]} (worst regret {pick[0]}, expected {pick[4]:.2f})")
if __name__ == "__main__":
main()$ python3 moat_futures.py # expected value and worst regret
moat A B C expected worst regret
verified environments at scale 8 7 8 7.70 0
customer skill library (today) 6 5 8 6.20 2
self-improving tool use 7 4 5 5.60 3
automated harness design 5 7 4 5.35 4
long-horizon agents 4 3 5 3.95 4
robust pick: verified environments at scale (worst regret 0, expected 7.70)
Scores in each future (0โ10)
Bar length = expected score; labels = score in each future. Real output.
๐ Go deeper: what "worst regret" means
- Regret is how much you lose by not having picked the best option for the future that actually happens.
- Environments are never beaten in any future, so their worst regret is 0.
- Choosing by worst regret does not need the probabilities, which is useful when nobody trusts them.
4. How long would a rival need?
copy_time.py ยท Python 3, no dependencies; scenario estimates
"""copy_time.py โ how long would a well-funded rival need to copy each moat?
months = build_months / (1 + rival_speedup) + data_months, where data_months
is the time to collect what cannot be bought (customer tasks, gate history).
All inputs are the story's SCENARIO estimates; the rival is assumed to have
3x Maestro's engineers (speedup 2).
"""
MOATS = { # build months for Maestro, months of data that cannot be bought
"verified environments at scale": (18, 12),
"customer skill library (today)": (12, 18),
"automated harness design": (9, 0),
"self-improving tool use": (12, 3),
"long-horizon agents": (15, 0),
}
SPEEDUP = 2
def main():
for m, (build, data) in sorted(MOATS.items(), key=lambda kv: -(kv[1][0] / (1 + SPEEDUP) + kv[1][1])):
months = build / (1 + SPEEDUP) + data
print(f"{m:32} {months:5.1f} months {'durable' if months >= 12 else 'copyable'}")
if __name__ == "__main__":
main()$ python3 copy_time.py # rival with 3x the engineers
customer skill library (today) 22.0 months durable
verified environments at scale 18.0 months durable
self-improving tool use 7.0 months copyable
long-horizon agents 5.0 months copyable
automated harness design 3.0 months copyable
Months for a 3ร bigger rival to copy
Real output. Engineering can be bought; customer data and history cannot.
- The lesson matches the startup notes.
- When intelligence gets cheap, value moves to what it cannot make by itself: data, trust and history (owner's notes, Sosin and Rauch).
5. Sandbox: bet on a future
๐ฎ Sandbox: change what you believe about the future
Same scores as moat_futures.py; scenario estimates.
- Try this.
- Push Future B (cheap open models) to 100 and the others to 0: harness design ties environments at 7, but it is copyable in 3 months.
- Push Future A to 100: tool use comes close (7 against 8).
- Whatever you set, environments are never beaten, which is why the robust pick stays the same.
6. Five mistakes
- Picking the moat that wins the single most likely future.
- Confusing "exciting" with "defensible".
- Ignoring how long a bigger rival needs to copy it.
- Treating the current moat as permanent.
- Never testing the exciting idea, so it keeps coming back as a distraction.
7. Delegation box
๐ Gopala himself
- The three futures and the final pick
- Telling the board what is not being bet on
๐ค AI agents
- Scoring sheets and scripts
- Competitor tracking
๐ค Bought, rented or borrowed
- Advisers for the futures workshop
8. Budget and risk tracker
| Item | Status after Part 7 | Note |
|---|---|---|
| Next moat | Environments at scale | Regret 0, 18 months to copy |
| Current moat | Skill library kept | 22 months to copy |
| Forecast risk | Watch | Futures are estimates |
| Focus | Good | Long-horizon gets a bounded test only |
9. The series
The Maestro, Season 2: 23 parts in five phases (Season 1: 18 parts)
VerdictThe next moat is the one that holds up whichever future arrives and that a richer rival cannot buy its way into quickly: verified environments at scale, alongside the skill library. Next: building them.
10. Reading the numbers
- Primary / standard
- Minimum-regret choice is standard decision analysis; the "complements" argument is from the owner's startup notes.
- Secondary โ ๏ธ
- None in this part.
- Our assumptions
- Futures, probabilities, scores and copy times are the story's scenario.