From July 2 through August 1, 2026, one primary account carried 3,913 merged public PRs while one human directed Agents across the Kungfu system. The complete public system recorded 4,065 merged PRs across 16 repositories.
In the same fixed 30-day window, all four visible author accounts in Google AX merged 52 PRs. The primary Kungfu account therefore carried 75.25x as many merged public PRs as the complete visible AX team.
OpenAI provides a second reference point. Its Harness engineering article reports roughly 1,500 opened and merged PRs over five months with a small team of three engineers driving Codex; the team had grown to seven by publication. Kungfu's primary-account count in 30 days is about 2.6x that reported five-month total. The complete Kungfu count is about 2.7x.
This report publishes the records, collector, checks, and source boundaries behind those observations. The interpretation boundaries are collected at the end so the evidence can be read first.
Explore the interactive charts, downloadable records, and machine-readable analysis at
[libkungfu.dev/dogfood/parallel-runtime-paths].
The fixed-window calculation is:
2026-07-02 00:00 UTC through 2026-08-01 00:00 UTC
3,913 merged public PRs under Kungfu's primary accountable account
52 merged public PRs from all four visible Google AX accounts combined
3,913 / 52 = 75.25
Rounded to the nearest whole number, the same-window Google AX comparison is 75x.
| System | Evidence | Observation window | Responsibility surface | Merged PRs |
|---|---|---|---|---|
| Google AX | Reproducible public GitHub records | 2026-07-02 through 2026-08-01 |
||
| One public repository; four visible author accounts | 52 | |||
| OpenAI Harness engineering | Official OpenAI first-party report | About five months after the first commit in late August 2025 | One non-public repository; three engineers driving Codex, seven by publication | About 1,500 |
| Kungfu Systems | Reproducible public GitHub records | 2026-07-02 through 2026-08-01 |
||
| Sixteen public repositories; one primary account carried 96.26% | 4,065 total; 3,913 primary |
The Google coordinate answers a strict same-window public-data question. The OpenAI coordinate answers a different question: how the Kungfu record compares with one of the strongest publicly described examples of an Agent-first engineering organization.
The scopes preserve the organization forms that produced each observation:
Google AX: one professional Agent-runtime repository with every visible PR author account included.Kungfu: one human directing Agents across the complete public product, protocol, build, release, site, and publication system. One GitHub account is responsible for more than 96% of its merged PRs in the operating window.OpenAI Harness engineering: a non-public, single-repository product built under a no-manually-written-code constraint. OpenAI reports that three engineers drove Codex, that the team had grown to seven by publication, and that roughly 1,500 PRs were opened and merged over five months.
Reducing Kungfu to one repository would omit the release and maintenance work that the same human-Agent system had to carry. Reducing AX to its leading author would overstate the contrast. The headline therefore compares one primary Kungfu responsibility identity with all visible AX responsibility identities combined.
The first strict 30-day window runs from 2026-06-16T00:00:00Z
through
2026-07-16T00:00:00Z
.
| Public observation | Google AX | Kungfu Systems |
|---|---|---|
| Merged PRs | 102 | 2,323 |
| PR author accounts | 5 | 3 |
| Active repositories | 1 | 11 |
| Active merge days | 21 | 18 |
| Median gross changed lines per PR | 141 | 124 |
Kungfu v4 engineering began on June 16, but systematic PR-mediated settlement did not begin until June 29. The record contains 99 default-branch commits over 12 active days before that boundary. Those commits remain commits rather than synthetic PRs. This window preserves the bootstrap history; the operating window below supplies the headline comparison.
The second strict 30-day window runs from 2026-07-02T00:00:00Z
through
2026-08-01T00:00:00Z
, after PR-mediated settlement is already operating.
| Public observation | Google AX | Kungfu Systems |
|---|---|---|
| Merged PRs | 52 | 4,065 |
| PR author accounts | 4 | 4 |
| Active repositories | 1 | 16 |
| Active merge days | 15 | 30 |
| Median gross changed lines per PR | 92 | 106 |
Within the Kungfu total, the primary account carried 3,913 merged PRs, or 96.26% of the observed total. Its median gross change size was 119 lines per PR, compared with 92 lines across all AX PRs. The count gap therefore coexists with a higher Kungfu median gross change size in this window.
Two ratios answer different questions:
4,065 / 52 = 78.17
: complete public Kungfu system versus complete visible AX team.3,913 / 52 = 75.25
: one primary Kungfu responsibility account versus the complete visible AX team. This is the more conservative headline ratio.
Agent participation is directly visible or explicitly reported in all three systems.
The AX default-branch history before the cutoff contains seven commits with an explicit Gemini Code Assist co-author marker. The Kungfu v4 history from June 16 through August 1 contains 135 uniquely attributed commits: 78 Codex, 34 Claude, 21 Cursor, and 2 Amp.
These markers make Agent participation publicly observable in both histories.
OpenAI's evidence is explicit but different: the company states that Codex wrote every line in the repository and that humans never directly contributed code. Its article also reports roughly one million lines and hundreds of internal users.
The exact public-data comparison shows one accountable Kungfu identity carrying a radically larger visible engineering and settlement surface than the complete visible Google AX team in the same 30-day window.
The OpenAI report supplies a second scale coordinate: Kungfu's primary-account count in one month exceeds OpenAI's reported total for five months. Together, the two references establish that the Kungfu record is unusual enough to demand an explanation of the operating system behind it.
The collector uses these exact GitHub Search scopes:
repo:google/ax is:pr is:merged merged:2026-06-16T00:00:00Z..2026-07-15T23:59:59Z
org:kungfu-systems is:public is:pr is:merged merged:2026-06-16T00:00:00Z..2026-07-15T23:59:59Z
repo:google/ax is:pr is:merged merged:2026-07-02T00:00:00Z..2026-07-31T23:59:59Z
org:kungfu-systems is:public is:pr is:merged merged:2026-07-02T00:00:00Z..2026-07-31T23:59:59Z
From this repository, run:
node scripts/collect-agent-output-comparison.mjs --window bootstrap
node scripts/collect-agent-output-comparison.mjs --window operating
node scripts/check-agent-output-comparison.mjs
node scripts/check-agent-output-comparison.mjs \
src/fixtures/agent-output-comparison-operating.snapshot.json
The checked-in snapshots are bound by these SHA-256 digests:
d355600f62371bec3b87e76974d0c449405037c8a50552f34c14a9ee5229ef29 agent-output-comparison.snapshot.json
abb6711e92f731e03fab3bca1bad5cdb686fa1d3295b42bff956d2dddba7c58e agent-output-comparison-operating.snapshot.json
Repository evidence:
Collector sourceComparison checksBootstrap snapshotOperating snapshotEditorial framing and pinned source authorities
Public evidence:
OpenAI: Harness engineeringInteractive comparison and chartsMachine-readable analysisComplete bootstrap recordsComplete operating recordsBootstrap SHA-256Operating SHA-256
The records are published so that readers can reject the headline metric, define a different unit, remove categories they consider inappropriate, or test whether another explanation fits the same public history better.
This is first-party analysis published by the Kungfu project. These boundaries apply to every number and comparison above:
- A merged PR is a public work item, not a feature, quality, maturity, labor, productivity, or value unit.
- Google AX and Kungfu use reproducible public GitHub records. OpenAI's figures are approximate first-party disclosures from a non-public repository; the linked OpenAI article is their authority.
- The OpenAI repository's exact PR records, dates, author distribution, and
matching 30-day slice are unavailable. The
2.6x
and2.7x
figures are raw contextual ratios rather than normalized productivity measures. - Google internal work and every other form of non-public work are outside the reproducible dataset.
- AX is measured as one public product repository. Kungfu is measured as the 16-repository public system carried by the same operating organization.
- GitHub author accounts are public responsibility identities. They do not establish employment roles, labor hours, or individual authorship of every line.
- Kungfu PRs act as settlement objects, while AX PRs follow a conventional contribution workflow. Review, merge, and PR-splitting disciplines differ.
- Agent markers establish observable participation in part of the AX and Kungfu histories. They do not establish Agent authorship of every change, equal Agent usage, autonomous operation, or a single cause for the gap.
- Gross line and file counts can include generated files, vendored material, formatting, and repeated edits.
- OpenAI reports hundreds of internal users for its product. This report does not claim equivalent adoption, product maturity, or user value for Kungfu.
- The observations establish an unusual public throughput pattern. They do not establish that Kungfu is superior, that the work units are equivalent, or that any one mechanism caused the difference.