Агентская разработка vs вайб-кодинг: опыт Авито по контролю качества ИИ-кода

Avito published a breakdown on Habr in 2026 on how to maintain code quality when developing with AI agents: Andrey Brovko, head of testing at Avito Auto, showed that on the SWE-bench benchmark, models grew over three years from 2% of tasks solved in 2023 to over 80% in 2026, but generation speed alone does not guarantee quality.

How agentic development differs from vibe-coding

Agentic development differs from vibe-coding in its control over the process: with vibe-coding, the user simply asks the model to "do it" and gets an unpredictable result, while agentic engineering guides the agent through a methodology — specification, tests, review — making the result reproducible. As Andrey Brovko puts it, vibe-coding works for informal tasks, but only a structured approach can be scaled within a company.

"AI agents generate code in seconds, but the speed of code generation doesn't mean product quality.

If development processes are poorly structured, agents just deliver defects faster," — Andrey Brovko, head of testing at Avito Auto.

What a pilot with 100 engineers showed

A pilot with 100 engineers at Avito found no statistically significant increase in objective efficiency metrics, and quality metrics stayed at the same level — despite high subjective satisfaction among the team. Avito interprets this as the early phase of the DORA J-curve of learning: people first master the tool and learn to check the generated code, and only then does the useful effect arrive.

The payback economics here are modest. According to the article's calculation:

• SWE-bench: 2% of tasks solved in 2023 → over 70% in 2025 → over 80% in 2026
• Pilot — 100 engineers, objective metric growth is statistically insignificant
• Developer salary — $2500/month, AI assistant subscription — $100/month
• A 4% speedup is enough to break even — that's 6.4 hours per month
• Recommended model context window — at least 32K tokens

To support this, Avito rolled out three internal services in 2026: skills-hub (a skills marketplace with quality and security checks), mcp-hub (agent access to the codebase, documentation, and bug tracking), and AVIDA (Avito Developer Assistant) for the safe autonomous launch of agents on a schedule or on request.

Where to focus testing

The focus is shifting left — to the specification rather than the final code. Avito is building its process around Spec-Driven Development (SDD): first a structured specification is written with user stories and acceptance criteria, then tests are written following TDD (also reviewed by a human), and only after that comes implementation. Every acceptance criterion is linked to a specific test, giving 100% test coverage of requirements. SpecKit from GitHub, Superpowers, and OpenSpec are mentioned as ready-made frameworks.

The priority for manual activities is shifting toward what models still do poorly: reviewing the specification, assessing product and project risks, reviewing tests and code, and exploratory and UX testing. A separate risk is that models can "lie" — reporting success when only stubs actually passed. To counter this, Avito applies mutation testing and a human-in-the-loop/agent-in-the-loop combination, and tracks progress using DORA metrics — Deployment Frequency, Lead Time for Changes, MTTR, and Change Failure Rate.

What this means

Avito's conclusion is straightforward: with weak architecture, testing, and release processes, AI agents simply speed up the delivery of defects, while with strong ones, they deliver speed without losing stability. The company is betting not on the model itself, but on the processes built around it.

FAQ

How does agentic development differ from vibe-coding?

Vibe-coding is a request to a model without process control and with an unpredictable result. Agentic development guides the agent through a methodology (specification → tests → implementation with review at every step), which makes the result reproducible and suitable for scaling.

When does an AI assistant pay for itself for a team?

According to Avito's calculation, with a developer salary of $2500 per month and a subscription of $100 per month, speeding up work by just 4% — about 6.4 hours per month — is enough for the tool to pay for itself.

What is Spec-Driven Development?

It's an approach where a structured specification and tests (following TDD) are written first, and the code appears last. Every acceptance criterion is tied to a test, which ensures 100% test coverage of requirements.