Covering the whole SDLC with an agent
Павел Федотовский «Как закрывать весь SDLC с помощью агента»
B · Useful, limited A Yandex engineer who runs internal AI-tool adoption reports a concrete n=1 production case with cost and effort figures; practical and specific, but a single self-reported experiment.
Summary based on the full transcript·Original: about 39 minutes to watch
In short
A 39-minute talk at a DotNetRu community conference (uploaded 22 October 2026) by Pavel Fedotovsky, who promotes AI tools inside Yandex and builds internal skill benchmarks. In May 2026, newly rotated to a team whose infrastructure he did not know and barely knowing Go, he built a production microservice only through prompts: the agent wrote code, deployed it, created the database, set up logs and alerts, and fixed production errors. It runs for internal Yandex consumers. Code was the smaller part of the effort; building the environment around the agent took most of the time.
Key numbers
- ~USD 2,000cost of taking one internal microservice to production through prompts only
- several weekstime the experiment took, mostly spent on the environment around the agent
- ~10tickets filed with internal infrastructure teams to make the agent workable
- ~1 hourtime for internal Yandex tooling to generate an MCP server from a service's code
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- Result: an API with database, alerts and monitoring in production, after several weeks and about USD 2,000, with roughly ten tickets to internal infrastructure teams.
- Context: AGENTS.md as entry point, with explicit notes such as Yandex using its own version control system "Arc" instead of git; a docs folder with plans, local-run instructions, pitfalls, architecture and decisions. He admits he does not manually validate the generated AGENTS.md.
- Tools: browser control works but costs tokens and raises security issues; MCP gives controlled access to wiki, tracker, logs and deploys; internal tooling generates an MCP server from service code in about an hour. A Yandex utility installs a team's shared skills and MCP set from one text file.
- Checks: local run is the most important enabler; slow deploy-to-test loops become the visible bottleneck once code is cheap. Infrastructure as code lets the agent change infrastructure itself. Production is still checked by hand because of sensitive data.
- Engineer role: planning first (about five iterations to a usable plan), automate every check, a session-end hook makes the agent update docs. He no longer reads code; a show of hands found almost everyone in the room still does.
How they got it
A personal experiment inside Yandex, described from memory at a conference. No logs, cost breakdown or quality metrics are shown. The digest is based on cleaned auto-captions with partly garbled tool names.
Limits
- One engineer, one small microservice; the speaker says the approach fits microservices best and is harder for monoliths.
- Cost and duration are approximate and self-reported.
- Yandex has unusually mature internal tooling (MCP generation, shared skill installers), which a typical firm lacks.
Why it matters for companies
The case moves the bottleneck from coding to environment: access, local runs, infrastructure tickets and test loops. Firms that want agents to carry more of the delivery cycle should budget for that platform work first, since it decided whether one engineer could ship. The figure of about USD 2,000 and several weeks for a production service by a non-expert gives a rough planning anchor, with the caveat that production checks stayed manual.
Key passages in translation
Russian original
на это у меня ушло несколько недель и порядка 2.000 долларов а на то, чтобы со всем этим разобраться.
It took me several weeks and around 2,000 dollars to work all of this out.
Russian original
основные здесь проблемы были даже не в написании кода, это меньшая часть времени. Большая часть уходила на то, чтобы вот как раз-таки эту среду вокруг агента развернуть
The main problems were not even in writing code; that was the smaller part of the time. Most of it went into setting up the environment around the agent.
Russian original
И мне кажется, что самое важное здесь - это локальный запуск.
And I think the most important thing here is the local run.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com
- Tags
- Coding agents, Mcp