Kungfu UNGFU™Developer Platform

Back to public dogfoodtwo fixed P30D observations / collected 2026-08-12

One human. Agents. Two thirty-day windows.

Google AX provides a professional multi-contributor Agent-runtime baseline. Kungfu provides two observable operating states: a v4 bootstrap before standardized settlement, and a full Buildchain operating window.

The central question is not whether one project opened more pull requests. It is how much public engineering surface one accountable human could carry through Agents, and whether that capacity changed when delivery responsibility became systematic.

Download bootstrap records Download operating records Audit PR size and deterministic samples Open analysis JSON
75× merged public PRs · one primary Kungfu account versus all 4 visible AX accounts combined

Why the first chart looked empty

Engineering started on June 16. PR-mediated settlement started later.

The first window must preserve both facts. Counting only merged PRs makes the bootstrap work disappear, even though the public default branch records sustained activity.

Jun 16first public Kungfu v4 commit
99default-branch commits before June 29
34explicitly authored by Claude (Code)
Jun 29Buildchain repository and systematic PR settlement begin

Brief technical background

Why these two projects belong in the same frame

Both are infrastructure for Agent execution that must outlive one prompt or process. Both make durable state, recovery, resumability, and inspection part of the runtime rather than an afterthought. They are not equivalent products, but their goals are sufficiently isomorphic to make the organizational contrast meaningful.

Sources: google/ax · kungfu-systems/kungfu

The comparison unit

Two real organization forms, kept intact.

Reducing Kungfu to one repository would remove the release, protocol, build-image, site, and publication work that the same human-Agent system actually had to carry.

Google AX

Multi-contributor, single repository

The public windows contain five PR author accounts working in google/ax. Its three leading accounts carry most of the visible change, giving the comparison a strong, non-tail baseline.

Observed: 5 author accounts in bootstrap · 4 in operating window

Kungfu Systems

One human, Agents, multiple repositories

Keren declares product direction and engineering boundaries; Agents carry most implementation, testing, and integration work. One primary account remains responsible for more than 96% of the observed merged work across the public system.

Operating window: 16 repositories · 96.3% from dongkeren

Window 1 · v4 bootstrap · strict P30D

2026-06-16 00:00 UTC → 2026-07-16 00:00 UTC

This window contains two Kungfu operating modes, so its merged-PR ratio is a conservative whole-period observation rather than a steady-state productivity measure. AX remained continuously active: 101 main-branch commits across 21 days.

Google AX · single repository

102merged public PRs
PR author accounts
5
Active merge days
21
Gross changed lines
38,000
Changed files
541
Median changed lines / PR
141
PRs / active merge day
4.9

Kungfu Systems · complete public system

2,323merged public PRs
Active repositories
11
Active merge days
18
Gross changed lines
1,378,586
Changed files
23,395
Median changed lines / PR
124
PRs / active merge day
129.1

Window 2 · Buildchain operating · strict P30D

2026-07-02 00:00 UTC → 2026-08-01 00:00 UTC

This second window begins after Buildchain and PR-mediated settlement are already operating. Kungfu records merged work on every one of the 30 calendar days.

Google AX · complete visible team

52merged public PRs
PR author accounts
4
Active merge days
15
Gross changed lines
8,384
Changed-file touches
194
Median changed lines / PR
92
Active repositories
1

Kungfu Systems · complete public system

4,065merged public settlements
Primary accountable account
3,913
Active merge days
30 / 30
Gross changed lines
3,698,331
Changed-file touches
53,052
Median changed lines / PR
106
Active repositories
16

Public output per accountable human

Compare one primary Kungfu account with the complete visible AX team.

This is stricter than comparing with an AX per-author average. It also avoids removing low-volume contributors: every AX author account in the operating window remains in the denominator.

75.25×merged public PRs: one Kungfu account versus all 4 AX accounts combined
3,913operating-window items under the primary Kungfu account
119 vs 92median gross changed lines per item; the count gap is not explained by tiny Kungfu PRs
96.3%of Kungfu operating-window PRs under one primary account

Measurement boundary: Kungfu PRs are settlement objects while AX PRs follow a conventional contribution workflow. These ratios measure visible responsibility throughput, not interchangeable features, labor hours, or code quality. Gross line and file totals remain in the raw dataset as stress tests, not as the headline efficiency coefficient.

Bootstrap-window merged PRs

The empty first half is a workflow boundary, not an absence of work.

The 99 pre-Buildchain commits above are intentionally not converted into synthetic PRs. The chart changes only when public PR settlement actually begins.

Google AXKungfu SystemsBars use one shared linear scale.
06-1670
06-1790
06-1860
06-1900
06-2000
06-2100
06-2230
06-2330
06-2420
06-2570
06-2690
06-2751
06-2800
06-29531
06-30273
07-016167
07-020118
07-03098
07-040139
07-050125
07-066202
07-076189
07-082129
07-098140
07-104195
07-111112
07-12092
07-133123
07-143163
07-155226

Buildchain operating-window merged work

Kungfu sustains public settlement on all 30 days.

The same linear scale is used for AX and Kungfu within this window. This chart shows operating regularity; it does not claim that one PR has identical semantics across the two systems.

Google AXKungfu SystemsBars use one shared linear scale.
07-020118
07-03098
07-040139
07-050125
07-066202
07-076189
07-082129
07-098140
07-104195
07-111112
07-12092
07-133123
07-143163
07-155226
07-16698
07-17377
07-18291
07-19092
07-20164
07-211130
07-220161
07-230144
07-240205
07-250153
07-260134
07-270119
07-281150
07-290119
07-300115
07-310162

Operating-window organization shape

Four AX contributor accounts. One dominant Kungfu account across sixteen repositories.

AX contributor identities are pseudonymized here because the comparison concerns organization form, not individual performance.

Kungfu repository distribution

  • kungfu-systems/buildchain1,451
  • kungfu-systems/kungfu1,333
  • kungfu-systems/build-images246
  • kungfu-systems/kfd232
  • kungfu-systems/site-libkungfu-dev171
  • kungfu-systems/site-kungfu-tech142
  • kungfu-systems/libnode94
  • kungfu-systems/paper-kungfu-product-white-paper87
  • kungfu-systems/paper-observer-declared-timelines65
  • kungfu-systems/runtime-images61
  • kungfu-systems/paper-kfd-foundation-real-world-agent-work60
  • kungfu-systems/agent-hub-demo54
  • kungfu-systems/paper-kfd-machine-life-roadmap47
  • kungfu-systems/homebrew-tap11
  • kungfu-systems/paper-episodes-to-primitives10
  • kungfu-systems/.github1

Google AX contributor accounts

  • AX contributor account A19
  • AX contributor account B16
  • AX contributor account C10
  • AX contributor account D7

Aliases follow merged PR rank in the operating window; accounts seen only in the bootstrap window are appended. Ties are resolved lexically. Public usernames remain in the downloadable records for reproducibility.

Kungfu PR authors

  • dongkeren3,913
  • app/kungfu-systems-release-bot85
  • kungfu-origin65
  • app/kungfu-paper-release-bot2

Explicit development attribution

Gemini on AX. Codex, Claude Code, Cursor, and Amp on Kungfu.

This evidence comes from commit provenance, not from product copy mentioning an Agent. AX records 7 default-branch commits explicitly co-authored by Gemini Code Assist before the comparison cutoff. Kungfu's public v4 branch records 135 unique explicitly attributed commits from June 16 through August 1.

Evidence boundary: AX's cited Gemini commits are repository-history evidence outside the two numeric windows. Kungfu's attribution census spans the v4 bootstrap and operating observations. These markers establish some Agent participation, not autonomous authorship of every change or Agent use as the sole cause of the output gap.

First-party interpretation

The anomaly is human leverage, not pull-request volume alone.

The public record does not prove labor productivity, feature value, or that Agent mediation alone caused the gap. It does show one accountable human carrying a radically larger visible engineering surface than the leading public contributors of a professional Agent-runtime project, with a second fixed window exposing the level sustained after Buildchain became operational.

Reproduce before believing

The conclusion is downstream of the records.

1. Two strict P30D windows

merged:2026-06-16T00:00:00Z..2026-07-15T23:59:59Z
merged:2026-07-02T00:00:00Z..2026-07-31T23:59:59Z

2. Complete scopes

repo:google/ax
org:kungfu-systems is:public

3. Full records

PR metadata, bodies, authors, timestamps, additions, deletions, and changed-file counts.

4. Coverage check

Per-repository records must equal independent GitHub Search totals; 1,000-result scopes are split by day.

5. Public collector

node scripts/collect-agent-output-comparison.mjs --window bootstrap
--window operating

6. Immutable digests

bootstrap:d355600f62371bec3b87e76974d0c449405037c8a50552f34c14a9ee5229ef29
operating:abb6711e92f731e03fab3bca1bad5cdb686fa1d3295b42bff956d2dddba7c58e

Download bootstrap records Download operating records Read collector source Verify SHA-256 Verify operating SHA-256

Comparison boundaries

What the public data does not prove

Exact public sources

Read both projects at the cited cuts.