Home โ€บ Blog โ€บ The Maestro, Season 2 ยท Part 9
The Maestro, Season 2 ยท Part 9 ยท Phase 6 ยท ๐Ÿ›‘ CEO hat

When Research Says No

The most exciting direction in Phase 6, agents that work for hours on one task, gets a fair test with written kill criteria. Six months later it fails three of them. This part closes Phase 6 with the kill itself: the criteria, the numbers, why the $1.9M already spent does not matter, what happens to the team and the knowledge, and why Gopala celebrates it.

FL
FrontierAGI Team
๐Ÿ’ผ CEO๐Ÿ”ฌ Research director๐Ÿง‘โ€๐Ÿซ Manager
โš ๏ธ How to read this (as of 3 Nov 2026) kill_decision.py and sunk_cost.py were run (real output below). The whole episode is the story's scenario: criteria, results, costs and probabilities are invented for teaching. Disclosure: Claude, made by Anthropic, helps write this site.

1. Mission log: months 27โ€“34

Week 108 ยท The test
  • Long-horizon agents get two researchers, six months and four written criteria.
Week 132 ยท Results
  • Three of four criteria fail (kill_decision.py).
Week 133 ยท The kill
  • One meeting; team moved to environments; write-up published internally.

2. The map: how a direction is stopped

๐Ÿ•ธ๏ธ From exciting idea to a clean kill

or click any node

Failed criteria in red, guards in green. Every number comes from the two scripts below.

3. The kill criteria at month 33 (scenario)

Scenario
  • Month 33: the long-horizon direction reports its results against the criteria agreed at month 27.
kill_decision.py ยท Python 3, no dependencies; scenario results
"""kill_decision.py โ€” the long-horizon agents direction against the kill
criteria written before it started (month 27). SCENARIO results, month 33.

A direction is killed if ANY pre-registered criterion fails at its checkpoint.
"""
CRITERIA = [  # name, target, observed, higher is better
    ("held-out tasks solved, 4h+ horizon (%)", 25, 11, True),
    ("gain over the L4 skill library baseline (pts)", 3.0, 0.4, True),
    ("cost per solved task vs product ($)", 2.0, 6.8, False),
    ("customers who would pay in interviews (of 12)", 4, 5, True),
]

def main():
    fails = 0
    for name, target, obs, up in CRITERIA:
        ok = obs >= target if up else obs <= target
        fails += not ok
        print(f"{name:48} target {target:>5}  observed {obs:>5}  {'pass' if ok else 'FAIL'}")
    print("decision:", "KILL the direction" if fails else "continue", f"({fails} of {len(CRITERIA)} criteria failed)")

if __name__ == "__main__":
    main()
$ python3 kill_decision.py # criteria fixed at month 27 held-out tasks solved, 4h+ horizon (%) target 25 observed 11 FAIL gain over the L4 skill library baseline (pts) target 3.0 observed 0.4 FAIL cost per solved task vs product ($) target 2.0 observed 6.8 FAIL customers who would pay in interviews (of 12) target 4 observed 5 pass decision: KILL the direction (3 of 4 criteria failed)

Observed vs target

4h+ tasks solved11% of 25% target
Gain over baseline0.4 of 3.0 pts
Cost per task (lower is better)target $2.0, observed $6.8
Would pay5 of 4 needed

Bar = observed as a share of target (cost: target รท observed). Real output.

4. Why the $1.9M does not count

sunk_cost.py ยท Python 3, no dependencies; scenario estimates
"""sunk_cost.py โ€” why the $1.9M already spent should not decide anything.

Compares the expected value of continuing six more months against stopping
and moving the team to the environments track. Probabilities and values are
the story's SCENARIO estimates; the money already spent appears in neither
option, because it is gone either way.
"""
SPENT = 1.9          # $M already spent (sunk)
CONT_COST = 2.4      # $M for six more months
P_SUCCESS = 0.12     # chance the direction recovers, given month-33 evidence
WIN_VALUE = 12.0     # $M value if it does
ENV_VALUE = 3.1      # $M expected value of moving the same people to environments

def main():
    cont = P_SUCCESS * WIN_VALUE - CONT_COST
    print(f"already spent (ignored): ${SPENT}M")
    print(f"continue: {P_SUCCESS:.0%} x ${WIN_VALUE}M - ${CONT_COST}M = ${cont:+.2f}M")
    print(f"stop and redeploy: ${ENV_VALUE:+.2f}M")
    print(f"break-even chance of success to justify continuing: {(CONT_COST + ENV_VALUE) / WIN_VALUE:.0%}")

if __name__ == "__main__":
    main()
$ python3 sunk_cost.py # compare only the future already spent (ignored): $1.9M continue: 12% x $12.0M - $2.4M = $-0.96M stop and redeploy: $+3.10M break-even chance of success to justify continuing: 46%

The only comparison that matters

Stop and redeploy to environments+$3.10M expected
Continue six more monthsโˆ’$0.96M expected

Real output. Continuing needs a 46% chance of success; the evidence says 12%.

๐Ÿ”Ž Go deeper: the sunk-cost trap, in one line
  • Money already spent is the same in both futures, so it cannot tell you which future is better.
    • Teams often keep a project alive "so the money was not wasted"; that spends more money to avoid admitting the first was lost.

5. Sandbox: try to save it

๐ŸŽฎ Sandbox: try to save the project

Same numbers as kill_decision.py and sunk_cost.py; scenario estimates.

6. What a good kill leaves behind

Three outputs of the kill

๐Ÿ—๏ธPeopletwo researchers join environments
๐Ÿ“„Knowledgewrite-up and three reusable parts
๐ŸŽ‰Culturethe kill is thanked in public

Story events.

7. Five mistakes

8. Delegation box

๐Ÿ™‹ Gopala himself
  • The kill decision and the all-hands message
๐Ÿค– AI agents
  • Assembling results against the criteria
  • Drafting the write-up for the team to check
๐Ÿค Bought, rented or borrowed
  • The independent director as witness to the criteria

9. Budget and risk tracker

ItemStatus after Part 9Note
Long-horizon agentsKilled at month 333 of 4 criteria failed
Money$1.9M spent, $2.4M savedFrom sunk_cost.py
Research engineHealthyExplores, kills, ships via the gate
MoraleWatchKill celebrated; team redeployed

10. The series

The Maestro, Season 2: 23 parts in five phases (Season 1: 18 parts)

VerdictPhase 6 ends with a "no": the most exciting direction fails its own criteria and is stopped in one meeting, without counting sunk money and without losing the people or the knowledge. That is what a working research engine looks like. Next, Phase 7: sizing and building Maestro-2.

11. Reading the numbers

Primary / standard
Sunk-cost reasoning and pre-registered criteria are standard decision practice.
Secondary โš ๏ธ
None in this part.
Our assumptions
The direction, criteria, results, costs and probabilities are the story's scenario.