Agentic engineering
vibecoding.tech is run by a machine of agents.
6,003 commits in 85 days — live feed below.
Documentation
Journal
Frontend
Backend
Admin
CI/CD
The machine keeps working while you read
you have been on this page 00:00 · machine pace since July 1: a commit every ~20 min · next one at this pace in ~20 min
See for yourself: the machine's last 24 hours
Pipeline events are real, time is compressed: a day plays out in a minute.
This is the real pipeline over yesterday: the machine found, the editor picked, the stories went out to the site. Time is compressed. The work is real.
Between “found” and “on the site” stands an editor: most of the raw material goes to the trash, and every rejection has a recorded reason. That is how this entire site was built: all four hundred pages.
This is agentic development. The whole map.1
Each rectangle is a region of the repository; the area is counted from commits since July 2, 2026. A human writes the rules and keeps the final word; the agents execute, and every commit stays in git. Click a region: it has its own story inside.
And this is vibecoding. Two regions out of six.
What people usually call vibecoding lives in two regions: the frontend with the public pages and the backend with the pipeline and the database. Less than a third of the machine.
And it is not only code. It is the whole company.
The same commits, a second coloring. All of vibecoding.tech is run by the same machine of agents: strategy, product, marketing and operations are described in specs. There is more documentation here than code — 387 thousand lines of specs against 391 thousand lines of code, and the code only executes what is written. A setup like this is called a Spec-Driven Company.
A human writes the rules and accepts the work
Evgeny Shilov is the founder of vibecoding.tech. Since 2023 he has coded with AI every day: custom LLM systems for business, an open-source agent in the GitHub top, a coding-agent benchmark on arXiv. In this machine he writes the rules up front and accepts the work after the fact; everything else is done by agents.
Okay, cheaper. And faster?
Faster by 13 times6
Okay, faster. But isn't it all neuroslop?
The machine throws out …
Slop or not — judge for yourself. This page was built entirely by agents: from a line in chat to production.
In 1913 Ford's assembly line closed the triangle for the automobile. In 2026 agentic development closes it.
| machine | team | ||
|---|---|---|---|
| price | ~$590 /mo | $56,375 | ×96 |
| speed | 54 days | ≈2 years | ×13 |
| quality7 | through the gate and selection | review and QA | ×1 |
| total | ≈$1K | ≈$1.3M | ×1,270 |
- Commit zones. The machine assigns each commit a zone by the paths of its files: docs/ and agent rules are "Documentation"; journal.md files are "Journal"; page and component code is "Frontend"; the pipeline, the database and scripts are "Backend"; the control console is "Admin"; the rest, configs and build glue, is "CI/CD". A commit's type is the dominant zone of its files. Merge commits are not counted.
- Commits. Counted on the main branch of the vibecoding.tech repository since July 1, 2026; the series is recalculated by the machine automatically, once an hour. The numbers are never edited by hand.
- Team equivalent. Replacement cost: the roles it would take to do the same work by hand, × the full cost of an employee to the employer (revision 28.08: median take-home pay from Habr Career H2 2025 × ~1.5 — income tax, social contributions ~30% up to the cap, hardware and hiring; senior level, judged by the machine's actual output). This is the price of people, not "value created": the headcount and the rates are a hypothesis (estimate 26.08) and our floor — every rate and count is editable right here in this section, and the total and the multipliers in the heading are recalculated honestly. The key metric for a role is a fact: what the machine already ships in that role, from the live snapshots on this same page. The live "while you're on the page" counter spends the payroll by the second: a month is taken as 30.44 days.
- Selection funnel. The total over the whole history of the store: counting functions page through the decisions of both editors (the X feed and sources v2) and the story store; a daily cron writes the sums, the page reads the ready snapshot. "Served" means the story went out to the site or to the Telegram channel.
- Machine costs. Subscription receipts from the mail collector + the OpenAI portal; token consumption comes from account status snapshots (the tokmax measurement); the $ rate is from the Central Bank of Russia, once a day. The page shows only the vibecoding.tech share, attributed by the paths of work sessions; Convex and the cloud are left out, people would need them too. The struck-through price tag in the duel and the second step of the ladder are the same consumed compute at public API prices: the build consumed ~$27,500 worth, paid for with ~$1,250 of subscriptions — the arbitrage, "knowing how to squeeze subscription compute", is itself part of the method; the numbers follow the "subscriptions" and "rate" knobs in this section.
- Speed. The delivery unit is a task "from the owner's message in chat to a merge into production"; the series is assembled from the chat history of work sessions × git, 115 tasks, snapshot 27.08. The scale is logarithmic, and the tail on the right is made up of long review sessions with the owner; autonomous pipeline branches (the feed) are not counted, they have no message behind them. Releases per day are counted from GitHub, 300 merges in 11 days, snapshot 26.08. The machine's days in the duel run by the build calendar since July 1. Sources for the "same work by a team" conversion multipliers: execution — our own series (the median machine task in minutes, by hand in hours; ×2 is ultra-conservative); focus — RescueTime measurements (~2 h 48 m productive per day) and Microsoft "Good Day" (~2.5 h of code); coordination — Brooks's law, n·(n−1)/2 channels, an hour per colleague per week = 2.5% of working time; dependencies — team flow efficiency is usually 5–15% (Vacanti, kanban measurements); absences — the Russian Labor Code, 28 days of vacation, plus holidays, sick leave and downtime while people are replaced. The formula is generous to the team on top of that: hiring and onboarding (+2–3 months to the first commit) are left out, as is the slowdown as the codebase grows. With the machine, written specs do the opposite: they make every next run faster.
- Quality. Neither machines nor teams have a single number for quality, so what gets compared is the release control loop. Nothing leaves the machine without passing selection: code goes through the gate (2,685 tests on every push + review), content goes through the selection feed (24,597 found, 3,292 served, 87% binned), its own ideas go through the decision log (1,268 of them, each with a "Rejected" field). The page does not award itself a quality multiplier — the quality row holds a live 0 to 1 knob, "does the machine's output measure up to a team" (what is being rated is this very page), and the summary total is honestly multiplied by your score: even at 0.1 it stays three-digit. The first rating you dial in goes into the hall counter (the median is shown from 10 ratings on); the counter is live, stored without personal data, and the audience is self-selected.