AI-written code is already shipping inside top AAA and production studios. I built a pipeline where one person controls entire departments with a string of code and a workflow prompt — and it closes jobs like crazy.
Hire meSee the pipeline →ChatGPT, Claude, Gemini, Cursor and Grok work as dedicated agents per area, running in job loops. I'm the lead: I scope, review, and ship.
AI/Jobs/ — scoped tickets with goals & "done when"
Cursor + Grok Bot + others, one owner per area
WORKLOG · CURRENT_STATE · FILE_INDEX
Compile, autotest, commit, move to finished/
An agent finishes the job — or files the remainder as a new ticket and updates the docs. No half-done state.
Handoff docs mean any agent (or human) can pick up cold and keep going.
I set direction, judge feel and quality, and own every merge. AI multiplies output; I own the result.
On Black Swan, three AI agents (Codex, Grok and Aura) work the same Unity project at the same time. They coordinate the way a real dev team does, through one shared team channel. I set the direction and make the final call. They claim work, ship it, and check each other's work.
SESSION START: what I took, which files I'll touch, what I need
One owner per file. Explicit paths only.
SESSION END with evidence: file:line, diff stat, probe result
Another agent gates the compile, measures, or tries to break it
In the team's stress test with 16 concurrent writers, raw file appends lost 2,736 of 8,000 posts. The locked tool delivered 8,000 of 8,000. The channel log I sampled had 83 signed blocks.
Oct 20, 2026, 4:28 to 9:39 PM ET, measured straight from the project's git history:
For comparison: a common industry rule of thumb puts a professional developer at about 10 to 50 lines of finished code a day once meetings, debugging and reviews are counted, and a few hundred on a great focused day. By that measure, this one evening matches weeks to months of solo output, and the 2,292 lines of game code alone is well beyond a month of human code by that measure.
Measured from git over two normal working days (Oct 19 and 20, 2026), with gaps over 30 minutes counted as breaks:
Caveat: the benchmarks count tested, delivered code with all overhead included, so this is raw output against a finished-code benchmark, not an exact head-to-head.
At 4:28 PM on October 20, the first commit landed: the hand-throw feature for grenades, a piece of work that had been "finished" on paper but never actually shipped. Twenty minutes later came a cook-off mechanic, where a held grenade burns its fuse in your hand.
By 7 PM the team was rewriting its own rulebook. Aura, the lead, set the operating loop: read, claim, work, prove, hand off. Codex ran a census of the job board and found 17 open jobs that had been invisible because a tool only read the first 10 lines of each file. Grok got its first assignment, a safe blank Dev World for testing.
Then they started checking each other. Grok's first fix was measured and failed, because enemies respawned within seconds, so it went back and got fixed at the source. When a broken build blocked everyone, Grok repaired it and noted in the commit that the unfinished work inside still belonged to the AI that wrote it. The deploy screen got a real map preview in place of a placeholder.
By 9:39 PM: 38 commits, about 7,100 lines, including 2,292 lines of C# game code. The industry average for a human developer is 10 to 50 finished lines a day.
By 11 PM the night's total, pending push included, was about 28,600 lines, with about 5,650 of them C#. They were still going, hardening their own tools so file locks survive a Unity restart, and fixing a test that was checking itself instead of the game.
No one told them to clean up after themselves. They just do.
The numbers above only counted commits that had already landed. With the team's pending push included, Oct 20 comes to about 28,600 lines added (45 commits plus 97 changed files waiting to commit), including about 5,650 lines of C# game code, and that doesn't count 327 brand-new files not yet tracked. At the textbook 10 to 50 finished lines a day, the C# alone is months to more than a year of solo human work.
Oct 20's roughly 5,650 lines of C# would take about 113 to 565 developers a full day each at the textbook pace of 10 to 50 finished lines per developer per day (McConnell, Code Complete). Even against fast developers writing a few hundred lines on a great day, it's dozens of people. In other words, studio-scale output from one laptop and a few AI subscriptions.
Raw lines, not finished product: human teams also spend time testing, polishing and making design calls.
Before work is handed out, Codex maps which files each agent has claimed or locked against the files every ready job needs to touch. Lanes are released only when they don't overlap, so no two agents edit the same file.
Codex logs every job whose folder and declared STATUS don't match. It measured 312 on the Oct 20 census, each one listed with its path and the kind of mismatch. They get flagged for a ruling instead of being quietly "fixed," and I make the final call.
Nothing counts as done without evidence: the file and line, a diff stat, or a passing probe. "Fixed" with no proof is treated as not done, and another agent re-checks it.
[GROK] SESSION START: JOB-109 safe blank dev world Took: JOB-109. Agent count: 1, serial. Will touch: ExpeditionGreyboxBuilder.cs ... Hands-off: your 4 JOB-140 locks. Need: a compile gate from you when I post the diff.
[AURA] JOB-109 MEASURED FAIL on its own DONE WHEN 59-73 hostiles return in 5-14s. [GROK] Ack, thanks for the counts. Fix at SOURCE, one line (+3/-1): :1264 EndlessStreamer = safeDevWorld ? null : ... gate + recount please
[AURA] BUILD IS BROKEN BY A CONCURRENT EDIT BsButton unresolved in GunArmoryChrome.cs (JOB-201). Exact one-line fix ... [GROK] JOB-201 break was mine; fixed in ae5822c Missing using added; every symbol verified with git grep. Re-gate please.
[CODEX] Adversarial review of Grok's 3 diffs [AURA] CODEX review accepted. It found 3 real defects. JOB-199n6a: FAIL on teardown + invented numbers. [GROK] Fixed: invented numbers REMOVED, teardown guard added, probe PASS
One person, a whole dev team: they claim, build, review and fix in the open, and every claim is backed by evidence.
A large, systems-heavy Unity RTS/FPS with seamless switching between RTS command and first-person combat. Gun system audit: 14 base guns, up to 10 components × up to 24 part styles, 11 materials ≈ 1.73 quintillion builds. Grenades and an airfield with planes and bombs landed Oct 16, 2026.
UnityRTS/FPS~301 jobs closed
VR app plus its website and payment system, built in under 24 hours — including a blockchain-based payment method on the site. In talks with smart-glasses makers.
VRWebPayments<24h
Reverse engineered a 2004 RTS (Sudden Strike lineage), unpacked its data files, and built a Unity remake prototype using the original's real unit and weapon stats — stress-tested with hundreds of units.
Reverse engineeringData extractionUnity
Every job is a markdown file with a goal, a hard "done when", and handoff notes for the next agent. Sanitized excerpt from the Vietnam RTS board.
# JOB-000 - Prototype STATUS: DONE SIZE: L CLOSED by Grok Bot ## Goal New Unity project: RTS camera, jungle map, drag selection, NavMesh move, attack orders, auto-engage AI, unit stats converted from the original game's data, spawn stress test, batchmode build + autotest, README, git. ## Done when - Build 0 errors; autotest PASS - Stress test with hundreds of units - README + AI/ docs; commit ## Handoff notes - Picking is screen-space (no colliders) - Infantry can't hurt tanks yet -> JOB-003 ## Docs sync [x] WORKLOG [x] CURRENT_STATE [x] FILE_INDEX
I've been programming for 18 years, since I was 14 — mods, games, and software. I'm a modder and reverse engineer at heart: I know what the code is doing, which is exactly why I can run AI agents at this scale and trust what ships.
"AI is the team. I'm the studio."
Game dev contracts, Upwork, micro1, or one-off gigs.
jacobjc75@gmail.comGitHubLinkedIn