On March 30, 2026, I used Codex inside my own governance framework, Virtual Company Kernel, to deliver and merge six substantial EPICs in roughly 3.5 hours. Based on the repo’s own role decomposition, that scope maps to about 457 hours of senior team-equivalent labor, or €13,138.75 at €28.75 per hour.
There are a lot of vague claims in AI right now. This is not one of them.
This post is about a bounded, auditable result: between 18:11 CEST and 21:45 CEST on March 30, 2026, Codex helped me deliver and merge EPIC-04 through EPIC-09 in the Virtual Company Kernel repository.
That scope covered:
- compatibility adapter core
- compatibility import and reconciliation
- guided starter bootstrap
- reference companion package
- evidence automation MVP
- ORCH control plane
Across the six feature commits behind that work, the repo shows 173 file-change entries, 8,203 insertions, and 117 deletions. More importantly, the output was not just code generation. It included analysis, architecture decisions, product framing, implementation, validation, and release discipline.
What Actually Happened
The first feature commit in this run was EPIC-04 at 18:11 CEST.
The last merge into develop was EPIC-09 at 21:45 CEST.
Within that window, Codex helped push six EPICs from defined scope to merged delivery:
- EPIC-04 built the compatibility adapter core for Claude Code, GitHub Copilot, and Cursor.
- EPIC-05 added compatibility import and reconciliation for existing instruction surfaces.
- EPIC-06 delivered a guided starter bootstrap flow.
- EPIC-07 produced a reference companion package with a real downstream proof path.
- EPIC-08 turned governed evidence into PR and CI review surfaces.
- EPIC-09 added a read-only ORCH control-plane view over governed delivery state.
This is the difference between “AI wrote some code” and “AI participated in governed delivery.”
Proof from the Repository
| EPIC | Feature commit | Merge commit | Scope |
|---|---|---|---|
| EPIC-04 | 159a481 | a12a31e | Compatibility adapter core |
| EPIC-05 | 35ac459 | b8d15d4 | Import and reconciliation |
| EPIC-06 | 8521fff | 8d40469 | Guided starter bootstrap |
| EPIC-07 | b03f862 | 5836ce4 | Reference companion package |
| EPIC-08 | a1988bf | c58f202 | Evidence automation MVP |
| EPIC-09 | 5fdc4fb | 6daa949 | ORCH control plane |
Human Team-Equivalent Estimate
To make the comparison concrete, I translated the completed scope into the same role model used by the repo itself: analysis, architecture, product, design, development, testing, and Git/release work.
This is not a claim that these were logged human timesheets. It is a modeled labor estimate derived from the repository’s EPIC and requirements-by-role breakdown for the delivered scope.
| Role | Hours | Man-days | Cost at €28.75/h |
|---|---|---|---|
| Analysis | 52 | 6.5 | €1,495.00 |
| Architecture | 45 | 5.625 | €1,293.75 |
| Product | 34 | 4.25 | €977.50 |
| Design | 6 | 0.75 | €172.50 |
| Development | 236 | 29.5 | €6,785.00 |
| Testing | 72 | 9.0 | €2,070.00 |
| Git / Release discipline | 12 | 1.5 | €345.00 |
| Total | 457 | 57.125 | €13,138.75 |
That is the number that matters.
It means one operator, with the right system around the model, compressed a very large amount of senior delivery labor into a very small time window.
Why Virtual Company Kernel Mattered So Much
If this story were only about Codex, it would be incomplete.
Codex was the execution engine, but Virtual Company Kernel was the operating system that made the output structured, reviewable, and mergeable.
Virtual Company Kernel played a decisive role in at least five ways:
1. It turned vague prompting into governed execution
The work was not “go build something useful.”
It was already decomposed into:
- EPICs
- role responsibilities
- wave sequencing
- gate rules
- validation expectations
- durable artefacts
That matters because AI performs far better when the problem is operationalized, not just described.
2. It forced source-of-truth discipline
Virtual Company Kernel defines what governs what:
- product truth
- operating truth
- workflow truth
- architecture truth
- audit truth
That prevents one of the biggest failure modes in AI-generated work: silent drift between intent, implementation, and documentation.
3. It made non-code work first-class
Most AI delivery demos collapse everything into code generation. Real delivery does not work like that.
The kernel forced the work to include:
- analysis contracts
- architecture boundaries
- product decisions
- validation evidence
- release and merge readiness
That is why the result can reasonably be compared to senior team-equivalent labor, not just programming hours.
4. It made validation part of the work, not an afterthought
A lot of AI-generated output looks fast until you ask whether it is testable, reviewable, or safe to merge.
Virtual Company Kernel bakes that in. The EPIC flow requires validation artefacts, linting, test execution, and explicit closeout. That means the result is not just “content produced quickly.” It is “delivery that survived proof.”
5. It created the evidence trail that makes this claim credible
The reason I can write this post with confidence is that the repo itself holds the proof:
- commit trail
- EPIC decomposition
- requirements-by-role breakdown
- delivered artefacts
- validation outputs
- merge records
Without the kernel, I might still have gotten fast output from Codex.
With the kernel, I got auditable delivery.
What This Post Is Not Claiming
To keep the claim honest:
- I am not saying a human team would necessarily need the same calendar time. Teams can parallelize.
- I am not saying the €13,138.75 is a literal payroll saving.
- I am not saying Codex acted alone in a vacuum.
I am saying this:
When you combine a strong execution model like Codex with a strong governance model like Virtual Company Kernel, you can compress a large amount of senior delivery labor into a remarkably small, verifiable window.
That is the real shift.
The Bigger Point
The interesting AI question is no longer just:
Can it generate code?
The more important question is:
How much senior delivery capacity can one person unlock when the model is wrapped in the right operating system?
On March 30, 2026, my answer was:
- about 3.5 hours of elapsed time
- 6 EPICs delivered and merged
- 457 hours of modeled senior team-equivalent labor
- €13,138.75 of equivalent work at the chosen rate
Codex made the speed possible.
Virtual Company Kernel made the speed trustworthy.