Nick Frolov

Great way to spot how model providers are cheating by memorising benchmarks datasets

Back in April I sat down with @iamitbeard (host of one of the largest Russian-speaking engineering podcasts). Almost 6 hours of non-stop talking later, I walked out feeling like I had completely overshared.

Today, the final 4-hour cut is live. While the interview itself is in Russian, the questions we unpacked are universal for anyone building software or leading engineering teams right now.

Instead of talking about AI hype, we went deep into the actual mechanics of Agentic SDLC, token economics, and how we build Junie at JetBrains.

Here is the breakdown of what we covered over those 4 hours:

My path from PHP developer and leading EPAM in the Netherlands to HFT, building our first GPT-3.5 prototype for traders, and co-building Refact.ai before joining JetBrains.

Where the real bottleneck in software engineering moved now that going from idea to prototype takes hours (and how JetBrains designers prototype directly with AI).

Spec-Driven Development: Where to store requirements and how to link specs, tests, and verification loops.

Why core engineering practices and tests matter more than ever when agents write the code.

Why JetBrains builds Junie when Claude Code and Codex already exist — and why we openly embrace 3rd-party agents in our IDEs via ACP ( plus an update on JetBrains Air and Junie since we recorded!).

If you speak Russian and want to watch links are below 👇

How to keep focus and make decisions when we have to manage dozens of parallel agent sessions? Decision-First UX is the most important problem to solve in this multithreaded world of agentic development.

Running 10 coding agents in parallel means 10 chat streams to monitor, 10 context switches and states to reconstruct, and 10 walls of text competing for attention.

The bottleneck moved from writing code to human attention and judgment—decision-making. Those are the real scarce resources for builders today.

Even in an automated workflow (research → spec → code → review → fix), every human step left in the loop is a decision:
Right or wrong? Keep going or roll back? Option A or B? Safe to merge?

Yet tools still force a slow, chat-first loop:
Open tab → read transcript → reconstruct state → inspect diffs → decide.

To optimize for decision speed and quality, we need four shifts:

1. Portable task cards, not chat tabs
Decouple the task from the session. On laptop or phone, a card shows at a glance: goal, state, changes, test evidence, and the decision needed next.

2. An exception inbox, not status noise
Let repair loops run unattended. Interrupt only on ambiguity, broken checks, or risk.

3. Batched judgment, not live babysitting
As DHH noted on Lex Fridman's podcast, "the human in the loop is the limit." Let agents validate work asynchronously and clear a batched digest of ready decisions on your own cadence.

4. Step-specific decision surfaces
Assemble context for each gate:
• Spec: assumptions, options, Approve/Redirect.
• Code: test results + one-click test stand.
• Review: risk map before raw diffs.

The winning metric isn't how many agents you spawn—it's time to the next right human decision.

Which of these four is the biggest gap in your workflows today?

Alibaba just hosted Aspara - a big AI Conference in Amsterdam. Showcased Qwen models and their agentic platforms. For me personally huge moment was seeing the Head of Qwen Models showcase JetBrains and Junie on the main stage.
Kudos to the team of Junie Local 🙌

I participated in the panel and discussed why it is hard to move from cool AI demos to enterprises scale.
With 4–5 new models releases a week enterprises need a stable platform layer which gives the freedom to switch between models and agents while keeping workflow stable and predictable. This is exactly what we are building with Air Platform.

Head of @Alibaba_Qwen showcase @junie_ai Local at Alibaba Conference