Skip to content
Thomas Fairhurst
Chapters

Chapter 12 · Subtraction

Delete what the loop doesn't use

Agents add machinery faster than anyone reads it, and every extra file is one more thing the next agent must get through. So I made deleting a discipline: whatever the working loop did not use went, about 353,000 lines in 15 hours, and review rounds went from 37 to none.

~353,000lines deleted in 15 hours, 5 to 6 Nov
37 → 0review rounds: 37 on one plan, at most 3 from 4 Nov, none by default from 5 Nov
3times the agents' main instruction file was grown, then cut back

Starting clean when a path goes nowhere

The first repo’s best learned bot won 12 of 140 games against the rules bot it was meant to replace. On 31 Oct I had everything saved, deleted it, and started a new repo with only what had proved its worth.

What the new repo took

Carried overLeft behind
The official game docs, unchangedThe orchestrator, its heartbeat and merge queue
The official engine and its native buildThe gates and the parameter tuner
70 maps, 25 edge-case mapsThe learned bot and its training code
v1 and every fifth version to v50, as opponentsOld commits, except as reference
The map, replay and fuel toolsEvery dashboard page but compute
About 165 GB of old runs deleted from the agents’ server and 74 GB from Imperial’s storage; 349 branches and a 5.4 GB release kept the rest. The server’s disk went from 92% to 53% full.

Proofs and plans: four days, then deleted

The new repo began with a plan I wanted proved: every step citing a register of axioms, backed by a Lean proof, reviewed until reviewers found nothing. Four days in, no version beat v50. Rigour about the plan was not progress on the bot, so I switched to building the bot.

  1. A register of axiomsAxioms, facts and decisions every step had to cite. 63 entries, then 85.
  2. Lean proofs and nonstop review“If we can formalize this somehow through some logic, maybe in Lean.”
  3. A second botv50 with every assumption written down as a cited “belief”.
  4. I see the plan has no end20 prerequisites before the loop could start. “It seems like it's just a never-ending path.”
  5. A new bot from first principlesThe lineage that shipped. Its first commit came 53 minutes later.
  6. The proofs and plans are deletedThe Lean proofs, the plan, the axioms and the second bot.

What the plan said of its own main theorem

today ε_cov has no useful bound, so x = 1 and the theorem holds trivially

A proof can be correct and still promise nothing. I had asked for the heaviest of this machinery myself, and I was the one who had it deleted.

Review rounds: 37, then 3, then none

The rule was fresh reviewers who did none of the work, round after round, until a round found nothing. Every round cost a fresh set of agents.

Highest review round reached, 31 Oct to 5 Nov

The second bot's loop plan
37
The register of axioms
31
Upload rules
28
Upload model
24

Agent threads started on 3 Nov, by job

  • WrappersHaiku, each running one review command
  • Workflow stepsOpus: fix, triage, record
  • Other Opus threadshigh or extra-high effort
  • ReviewersGPT 6.1
Of the 63.5 agent-hours spent inside commands that day, 52 went to waiting on reviewers. Of 28 re-reviews forced when a merge was rejected, 21 found nothing.

How I cut the rule back

  • Nonstop1 Nov
  • One reviewer, every lens3 Nov
  • At most 3 rounds4 Nov
  • None by default5 Nov
  • Only when I ask6 Nov

One night, 353,000 lines

On 6 Nov I asked for everything legacy to go. Most of it went between 02:00 and 05:00.

Lines deleted, by what they were

Lean proofsnothing the bot's loop ran read them
125,617
The second bot and its loopnever passed its gate against v50
92,330
The formal plan166 commits and its inputs
34,843
Hand-written engine portnever played a bot
12,384
Register of axioms85 entries
10,962
Work ledgerwhere sub-agents took their next task
9,297
Learned-bot specsthe reinforcement-learning route
6,587
Mutation tests64 to 65 lists
6,036

The mutation tests, planted bugs that check each test catches what it claims, show the trade most clearly.

What the mutation tests cost, and bought

1,311planted bugs in 33 lists, 3 Nov
576 of 4,023commits edited a list, dozens only to re-anchor text
~40commits fixed a weak test a mutant exposed

The bot that shipped grew up in the lighter repo: its rating went from 2,244 on 4 Nov to 2,660 at the ladder freeze.

Instructions grow; I cut them back

Most rules came from one incident each, so each repo’s main instruction file grew until agents got lost in it. Three times I had it cut back.

Root instruction file, in KB

“Leave no references to legacy stuff.”
6 Nov 02:55, before 44 instruction files became 22
“Set up the repo so that AGENTS.md is clean and clear.”
12 Nov 23:34

Keep the bench that pays

From 7 Nov agents built a scored bench for everything they wanted to climb. The ones that sped up tools paid; no bench of the bot’s play was ever shown to predict strength. On 11 Nov I kept one.

27benchesbuilt 7 to 11 Nov
0 of 4evals provena 9 Nov check of whether they predict strength
45,534lines deleted11 Nov, 03:58 to 07:24
1keptthe bot rebuild bench
The rented test machines went the same morning, at an estimated $8 to $18 a day. The test suite moved to pytest and ran in about 3.5 minutes.

The last loop goes too

My last orchestration loop had its own tools, dashboard and ledger of bets. When it started to get in the agents’ way, I had all of it deleted at 23:58 on 12 Nov.

110files deleted
11,618lines, the dashboard and the bets ledger included
4agent skills left at the end
“Delete everything related to the loop.”
12 Nov 23:34