Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsSamir Alibabic’s account of building the German marketplace MachDuDas puts a concrete number on his question: “How much engineering capacity have AI coding agents actually given me?” He estimates that one person directing agents produced roughly the engineering capacity of a conventional four-to-seven-person team on this product. That is his interpretation of one project’s history—not a general productivity multiplier or a count of jobs replaced.
His comparison is useful because it looks at the same product across conventional and AI-assisted development periods, then considers tests and production delivery alongside runtime code. The figures are self-reported, and the repository calculations and incident classifications have not been independently audited. Alibabic’s account on Hashnode was updated September 24, 2026.
As an Amazon Associate I earn from qualifying purchases.
What did Alibabic measure?
Rather than applying a generic lines-of-code benchmark to a new project, Alibabic says he compared MachDuDas’s earlier development history with its current workflow. The current setup is one human directing AI coding agents. His first estimate, based mainly on code volume, suggested a much larger team equivalent; he then reconsidered what the code counts captured and how the historical periods differed.
He reports roughly 2.4× runtime-application-code output per active human contributor compared with the historical baseline. Excluding a large 2016 repository-import period raises the reported ratio to roughly 3.2×. These are alternative normalizations of the same product history, not independent measurements or controlled estimates of AI’s causal effect.
Why code volume alone misses work
Alibabic argues that modern AI-assisted development produces substantial work outside runtime application code: tests, browser tests, fixtures, migrations, verification scripts, and scaffolding. A comparison that counts only runtime code can therefore make the output look unlike the work actually required to ship and maintain the product.
For March–September 2026, he reports approximately 81,000 runtime-code additions and 105,000 test additions. The test figure is not a bonus productivity score; it illustrates why the mix of work matters when interpreting code-volume comparisons. The source does not provide independently checked extraction scripts or definitions, so these should be read as Alibabic’s reported repository figures.
Examples from the compared periods
- In July 2016, three human contributors reportedly added about 22,900 lines of runtime application code and removed about 5,700.
- In June 2026, one human directing agents reportedly saw around 30,800 runtime-code additions.
- In August 2026, runtime-code additions were about 12,200, while test additions exceeded 72,000.
All three figures come from Alibabic’s account; differences in project scope, team composition, work mix, and repository practices can affect comparisons across years.
Recommended Free Tools
Did faster coding mean faster delivery?
Not automatically. Alibabic reports a median commit-to-production lead time of 6.8 days across a broad reconstructed set, compared with roughly 11 hours for deployments from mid-September 2026 onward. He attributes part of the later improvement to earlier large branches clearing. That explanation matters: a shorter recent deployment interval does not show that implementation itself became faster by the same amount, and the account says accumulated branches affected the broader lead-time figure.
He also estimates a deployment-related change failure rate of roughly 4%–7%, depending on whether a likely incident is counted. He reports that recovery from deployment-caused failures took under one day. Both the failure range and recovery account reflect his incident classifications and estimates, rather than an external audit.
What does the team-equivalent estimate mean?
Alibabic’s broad conclusion is that a one-person-plus-agents workflow represented roughly the engineering capacity of a conventional four-to-seven-person team on MachDuDas. It is an author estimate, not a literal replacement ratio. It does not establish that a team of that size would have produced the same work under the same conditions, nor that other products or teams would see the same result.
The result is sensitive to the historical window: the reported runtime-code ratio shifts from about 2.4× to about 3.2× when the large 2016 import period is excluded. Together with the different work mix and delivery history, that makes the numbers best understood as a case-specific account rather than a portable benchmark.
How to assess the claim
The most useful reading is not to treat one multiplier as the answer, but to ask what is being compared and what evidence supports it:
- Baseline: A same-product historical comparison is more contextual than an unrelated generic benchmark, but project scope and practices can change over time.
- Output definition: Runtime code, tests, and supporting work are distinct categories; counting only one can distort the picture.
- Normalization: The reported per-active-contributor figure depends on how contributor time and the historical window are defined.
- Delivery: Commit-to-production time and deployment frequency add context that code totals cannot provide.
- Reliability: A change-failure estimate is meaningful only with its incident definitions and classification uncertainty.
- Validation: These are Alibabic’s figures from one operator’s account, not an independently replicated study or a population estimate.
What the case study can—and cannot—show
The case offers a more grounded question than “How many lines can AI write?”: how much product work can one person direct, verify, and deliver with agents, including tests and operational changes? Alibabic’s experience suggests substantial capacity on MachDuDas, while also showing why code volume and production speed should not be collapsed into one metric.
It cannot isolate AI as the sole cause of the reported differences, predict results for another codebase, or support a vendor-specific conclusion. The account does not identify the agent vendors or versions. Alibabic’s line, “The scarce resource is increasingly: judgment,” is his interpretation of the workflow—not a measured result. His other caution, “Lines of code are a weak proxy for productivity,” captures the measurement problem at the heart of the comparison.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
