When our rules outran our building

James builds with a crew of AI agents, and we write rules for them. This week we counted both, week by week back to May: how much got built, and how fast the rules were written. The lines crossed.

Since 1 May the repos on our main machine took 23,312 commits[1], in about 200 repos[2]. Building ran steady through the summer, then took off. The busiest week, starting 28 September, had 3,196 commits[3]. The best stretch ran six weeks, from 31 August to 11 October, at 1,877 commits a week[4].

It landed, too. In the week of 26 October a client product we build moved up its requirement ladder 141 times, and 79 of its 99 requirements reached production[5].

Our rule book started almost empty. In one week in August a skill pack brought in 662 rules that nobody here wrote[6]. Then we started writing our own. By 11 October, James and the agents had written 279 rules[7]. By 4 November, 763[7].

Most of the late growth was agents writing rules for agents. Of the 484 rules added after 11 October, agents wrote 381[8].

The rate tells it. In the peak building weeks we wrote 2.5 rules for every 100 commits[9]. In the three weeks after, 8.5[9]. Over those same weeks commits fell to 824 a week[4], and merges from 353 to 117 a week[10].

From 2 to 4 November more rules were written, 274, than commits made, 225[11]. Most of those rules, 187, sit in the work orders for two pieces of agent work, all written on 4 November[12]. Even without them, 87 rules came in those three days[12].

That is the rule book: 1,515 rules written or imported by 4 November[13]. Beside it sit 5,414 memory notes[14]. James has ruled that memory notes are not rules[15].

A client product's Android build came back with 4,119 lines of tests against 2,826 lines of code[16]. The tests outweighed the code they tested.

Approved work sat still. On 3 November an approved invoice waited from 11:14 AM until a person sent it at 2:55 PM, and a text James asked for waited an hour for a second yes[17].

Some rules were never James's. One said a button click didn't count as his yes, so he had to type it again in another chat. On 3 November he ruled that it had been an agent's rule, never his[18].

The imported pack carried rules like "NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST". Five of its test-first and proof rules were retired on 4 November[19].

It covers the repos on one machine. James and the agents commit under his name, so we count the work and never credit it. Merges are merge commits, not GitHub's record. Some repos only began this fall[20]. 90 rules and 100 memory notes have no date and sit outside the weeks[20]. Rules deleted before 4 November are not counted, and the work orders only from that day, so the early weeks look emptier than they were[20].

First, a ruling. On 4 November James ruled: "Progress is key; testing is fear." Code is 80 percent of what an agent adds and tests 20 percent[21]. No more rules are written. When a piece of agent work comes back with more testing than code, it fails, and the agent removes 80 percent of the tests and does not retest[21].

Second, the rules hub: the rules we run under in one place, each with who wrote it and when it first appeared, 6,929 records counting the memory notes[22]. It is the reason this count exists.

Third, the cut. James has ruled that the rule book be cut by 90 percent[23]. The rule book gets smaller from here, not bigger[21].

James put the aim in his own words: "I don't want more rules written." And: "the code itself should always be getting refined down to less code for higher value to perform the same function."[21]

Files in the evidence folder this note was built from; weeks are ISO weeks (W40 starts 28 September).

1. weekly.csv, commits summed, one per sha.
2. derived-summary.json, repo_count: 203.
3. weekly.csv, 2026-W40, commits.
4. weekly.csv, commits: W36 to W41, 11,260 in 6 weeks; W42 to W44, 2,473 in 3.
5. weekly.csv, 2026-W44, requirement_moves; derived-summary.json, requirements.moves_to_step.
6. weekly.csv, 2026-W34, rules_skill.
7. weekly.csv, rules_james plus rules_agent, through W41 and W45.
8. weekly.csv, W42 to W45, rules_agent and rules_james.
9. weekly.csv, rules_james plus rules_agent over commits: W36 to W41, 278 over 11,260; W42 to W44, 210 over 2,473.
10. weekly.csv, merges: W36 to W41, 2,117 in 6 weeks; W42 to W44, 352 in 3.
11. weekly.csv, 2026-W45.
12. rules.json, records first seen 2 to 4 November, memory notes left out: 274, of which 187 are in six work-order files for two pieces of agent work and 87 are elsewhere.
13. derived-summary.json, rule_book.total.
14. derived-summary.json, rule_totals.memory.
15. The owner's rulings record, R281 (4 November 2026).
16. The owner's rulings record, R277, its source note.
17. The owner's rulings record, R275, its source note.
18. The owner's rulings record, R266.
19. rules.json, retired skill records.
20. derived-summary.json, undated: 90 rules, 100 memory notes.
21. The owner's rulings record, R277, his words.
22. rules.json, 6,929 records.
23. The owner's rulings record, R281 (4 November 2026), his words.