Home โ€บ Blog โ€บ The Maestro, Season 2 ยท Part 7
The Maestro, Season 2 ยท Part 7 ยท Phase 6 ยท ๐ŸŽฏ 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.

FL
FrontierAGI Team
๐Ÿ’ผ CEO๐Ÿ”ฌ Research director๐Ÿงญ Strategist
โš ๏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.

1. Mission log: months 30โ€“32

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

๐Ÿ•ธ๏ธ Which moat survives which future

or click any node

Futures in red, candidate moats in purple, tests in blue. Edge labels are scores (0โ€“10) or months to copy.

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 (0โ€“10)

Environments ยท A/B/C8 ยท 7 ยท 8
Skill library6 ยท 5 ยท 8
Tool use7 ยท 4 ยท 5
Harness design5 ยท 7 ยท 4
Long-horizon agents4 ยท 3 ยท 5

Bar length = expected score; labels = score in each future. Real output.

๐Ÿ”Ž 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

Customer skill library22 months
Environments at scale18
Self-improving tool use7
Long-horizon agents5
Harness design3

Real output. Engineering can be bought; customer data and history cannot.

5. Sandbox: bet on a future

๐ŸŽฎ Sandbox: change what you believe about the future

Same scores as moat_futures.py; scenario estimates.

6. Five mistakes

7. Delegation box

๐Ÿ™‹ Gopala himself
  • The three futures and the final pick
  • Telling the board what is not being bet on
๐Ÿค– AI agents
  • Scoring sheets and scripts
  • Competitor tracking
๐Ÿค Bought, rented or borrowed
  • Advisers for the futures workshop

8. Budget and risk tracker

ItemStatus after Part 7Note
Next moatEnvironments at scaleRegret 0, 18 months to copy
Current moatSkill library kept22 months to copy
Forecast riskWatchFutures are estimates
FocusGoodLong-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 โš ๏ธ
None in this part.
Our assumptions
Futures, probabilities, scores and copy times are the story's scenario.