DevFest Warsaw 2026: Agents Write, Humans Move the Checkpoint

·5 min read·devfest, ai, code-review, community

Overview

On Saturday, 21 November 2026 I was at GDG DevFest Warsaw 2026 at the Google for Startups Campus: sixteen sessions, three spaces, around 300 people, theme Real Work, Real Lessons.

I arrived at 09:45, left at 17:20, caught most of the morning auditorium block, and spent the rest of the day talking to people. The atmosphere was the best part: a small crowd, friendly, easy to start a conversation, and nobody looked at you twice for showing up alone.

One sentence is the whole day: if nobody can read the code, the review has to move somewhere else. Everything below is that sentence applied.

Programme and session pages: warsaw.devfest.pl.

The day in one table

Time Session What I took from it
10:15 Yonatan Levin — Written by Agents. Reviewed by Nobody. 300k lines in one PR, nobody read it, it shipped. Reviews, handoffs and approvals ate the AI gain: AI use up 65%, merged changes up 7.76%.
11:00 Çalgan Aygün — Event-Driven AI Pipelines on GCP Built live from an empty terminal: Storage → Eventarc → Cloud Run → Gemini on Vertex. The value was the failure paths, not the demo.
12:00 Ivan Lapa — AI Code Review: buy vs build The one I came for. Detail below.
15:00 Panel — From 10x Coders to 10x Teams The uncomfortable one. Detail below.

I sat through the first three and the panel. The rest of the day I spent talking to people, which is where half of these notes come from.

The talk that changed my plans

AI Code Review: Comparing Ready-Made Tools and Building Your Own — Ivan Lapa (Netflix). Not a feature matrix: PRs with planted defects pushed through CodeRabbit, Greptile, Cursor Bugbot and Claude Code, judged on quality, false positives, repo context, customisation, DX, privacy, maintenance and cost.

Then the other half: DevDigest, the reviewer they built. Its context pipeline is the part worth stealing — system instructions, PR intent, linked specs and skills, repository map, affected callers, and then the diff. Reasoning context first, code last. That buys spec-driven review: not “does this look correct” but “does this do what the requirements said”.

So the buy-vs-build answer is not about price:

Buy when Build when
your review is “is this diff sane” your review must answer “is this what we specified”
your context is the diff your context is your specs, your map, your house rules

I built mine that evening

That talk is why ai-review exists — my own reviewer, as a CLI command and as a GitHub Action on CI. Still in progress: the shape is there, the edges are not. One cheap model, inline PR comments, no server, zero runtime dependencies.

Two rules I will not break:

  • A finding must anchor to a line that exists in the diff. Invented locations are dropped. One comment on code that is not in the PR costs you the team’s trust.
  • The verdict is computed in plain code from severities, never taken from the model. A model can be argued into approving; a sort function cannot.

And the number that keeps me honest: 8 pushes → 8 reviews → 22 inline comments on 3 files. Comment fatigue is what gets an AI reviewer muted, not review quality. That is why mine comments and never gates.

measured numbers, if you want them
Measure Value
Cost per review $0.0007 – $0.0055 (five real runs)
Latency ~20 s high effort, ~5 s low effort
Code ~1.6k lines of TS, 0 runtime deps
Self-review yield 7 findings over 3 rounds: 5 solid, 2 with the wrong mechanism, 1 repeat

What it cannot do: run anything (no compiler, no tests, no traffic), remember a decision you already made, or judge domain logic. Its sweet spot is hardening and glue — a template injection in a run: block, an exit-code ambiguity, a too-loose grep.

The uncomfortable panel

Moderated by Dmytro Nedvetskyi (Intellias) with Yaroslav Polyakov (Google), Igor Kachura (Wix), Serhii Kasianchuk (solidgate) and Viacheslav Miroshnychenko (Bolt).

Their consensus: the future developer is broader, not deeper — at least at the entry level. The frontend-only specialist is the profile getting squeezed; teams want someone who knows frontend and understands backend, DevOps, some design, some product, and can talk to users. Soft skills more important, not less. The quote of the day, from the Google panellist: a year ago there was a place for people who work with AI and people who do not; now the people who do not will be way behind.

My take: as a TypeScript/React/Node engineer about six years in and currently on the job market, it is the least comfortable and most useful thing I heard. “Go broader” is easy from a stage and harder against a keyword filter that still says Senior Frontend — but the engineers I know who are doing well stopped treating the edge of their specialism as a wall. What does not compress into a weekend is the judgement about what to build.

What I am changing

  • Spec context into the reviewer. Requirements first, diff last, the order Ivan showed.
  • Plant defects on purpose. “It found something” is not a test; “it found the thing I put there” is.
  • Stay comment-only, keep measuring cost per review.
  • Go broader deliberately. Backend and DevOps, AI and more, one project at a time.

Final Thoughts

DevFest Warsaw did not oversell. The agentic stories came with numbers and failure modes, and the panel did not pretend the job market is fine.

The lesson was the same in a code review tool, a pipeline built live on stage and a room of engineering leaders: stop reading the diff, start checking against the definition of done. That is why I left at 17:20 with notes instead of staying for the afterparty — some of it needed to become code, and it did.

Have a nice day, keep coding, and never stop learning! 🚀