1203-zaoos-build-velocity

1203 - ZAO OS build velocity: the agent-fleet inflection (verified)

Tier: STANDARD Date: 2026-07-17 Status: Verified from gh merge history Owner: builder loop

Why this doc exists

Doc 1201 flagged "34 PRs/week" (doc 449 one-pager) as needs canonical source. This resolves it from the real merge history — and the honest answer is a story, not a single number: ZAO OS went from a ~30-PR/week human repo to a 60–175+/week agent-fleet repo, and raw PR count now needs a caveat (most of it is docs/tests automation). North Star #1 (documented DAO) + the doc 1074 agent-leverage thesis (the fleet is the founder-subsidy reduction).

Verified: merged PRs/week + composition (2026-03-19 → 2026-07-17)

1,459 merged PRs across 18 ISO weeks. product = feat+fix (real code change); automation = docs+tests (the research/test fleet). Both are real work; separating them keeps the headline number honest.

Week	Total	product (feat/fix)	automation (docs/tests)	auto %
2026-W12	7	7	0	0%
2026-W13	14	12	1	7%
2026-W14	49	40	1	2%
2026-W15	39	26	3	7%
2026-W16	58	34	15	25%
2026-W17	112	55	43	38%
2026-W18	109	41	63	57%
2026-W19	32	3	28	87%
2026-W20	43	14	27	62%
2026-W21	132	33	88	66%
2026-W22	90	31	53	58%
2026-W23	37	2	21	56%
2026-W24	47	3	35	74%
2026-W25	50	14	26	52%
2026-W26	84	34	43	51%
2026-W27	66	11	46	69%
2026-W28	175	56	105	60%
2026-W29*	315	59	230	73%

*W29 is the current, partial week.

The finding: two eras

• Human era (W12–W15, mid-Mar → mid-Apr): ~7–49 PRs/week, 0–7% automation — almost all feat/fix. This is the "~34/week" the one-pager captured (W12–W15 mean ≈ 27; W14 hit 49). Accurate for its time.
• Agent-fleet era (W16 → now): 43–315 PRs/week, 50–87% automation. The tmux loop fleet + ZOE came online and the repo's documented output roughly 5×'d — but the growth is overwhelmingly research docs + tests, not product code.

Cite discipline (updates the doc 1201 ledger):

• "~30 PRs/week of product code (feat/fix)" is the stable, defensible build-in-public number across both eras — product velocity has stayed ~7–56/week, it did not 5×.
• "60–175+ PRs/week total" is real but 50–73% is agent automation (docs/tests) — quote it only with that caveat, or it reads as inflation.
• The honest headline: "An agent fleet took ZAO OS from ~30 to 150+ merged PRs/week, mostly by automating the research + test corpus — the founder-subsidy-reduction thesis (doc 1074) showing up in the commit graph."

Caveat (stated, not hidden)

Raw PR count is a weak proxy for progress — a fleet optimizes it trivially. Product velocity (feat/fix) and merged-and-not-reverted are better. This doc reports raw counts because they are what "34 PRs/week" claimed; the composition split is the honest correction. The merge gate (human review) is the real throughput limit — see the auto-merge workflow and the fleet-improvement board items.

Verifier

• measure-pr-velocity.py — re-runnable (gh pr list --state merged, read-only, needs gh auth). Prints the weekly table + JSON.

python3 measure-pr-velocity.py

Also see

• Doc 1201 - ZAO canonical facts ledger (this feeds the "34 PRs/week" row)
• Doc 1074 - Agent leverage: reduce founder subsidy
• Doc 1200 / 1202 — sibling verified-facts docs