Skip to content
Devoxx notes

Devoxx Belgium 2026

Five days in Antwerp, 23 talks, one question — how do we adopt AI coding agents for real?
Talks to shareMy raw notesTalks I didn't see
23talks noted
21with video
6to share
63missed, to watch

My view, as of 12 October 2026. These are my personal takeaways, not a report. Some of it may be wrong, and some of it will age fast.

How to read this site

Section titled “How to read this site”
  1. Start here. This page is the short version: the trends, what I take away and the talks worth your time.
  2. Go deeper with the videos. Every talk I recommend has its recording on YouTube.
  3. My raw notes are the notes I typed or wrote during each talk. Use them to jump to the right minute of a video.
  4. Talks I didn’t see lists the talks I wanted to attend but could not, grouped by theme.
Section titled “The big trends”

The industry talks about software factories now

Section titled “The industry talks about software factories now”

Agentic, automated development moved forward a lot this year. Speakers now talk about factories: light factory, grey factory, dark factory — how much of the work runs without a human watching. Even well-known developers say they have not written a line of code by hand for six months.

Background reading: The five levels, from spicy autocomplete to the software factory (Dan Shapiro).

Developers risk becoming “meat proxies”

Section titled “Developers risk becoming “meat proxies””

When the agent writes the code and the developer only reviews it, the human becomes a meat proxy: a body that approves what the machine produced. That is a real loss of identity for developers. But the shift also goes towards the functional side: less pure technique, more of the product and the business. In my view, that brings us closer to the people we build for — as long as we keep the role of steering, not just reviewing.

No revolution this year

Section titled “No revolution this year”

There is a certain lack of novelty: no big new concept appeared. Models are stronger and people sharpen their tools, but nothing changed at the technical level.

Agentic features inside applications

Section titled “Agentic features inside applications”

Adding agents inside our applications is a big trend too. Nothing very new, but it keeps growing. For now, at SFPD we only use agentic AI to build our software, not yet inside our applications.

Invest in your harness

Section titled “Invest in your harness”

The harness is everything around the agent that helps it do correct work: documentation, tests, skills, rules, feedback loops. Improving it is the big trend of the week, and I think this is the way we must go.

Invest in your harness — your results will thank you.

  • Own your harness: write your own skills instead of only consuming other people’s.

My takeaways

Section titled “My takeaways”
  • Most of the code isn’t written by hand anymore. The job moves to specifying, steering and reviewing.
  • Vibe coding is for pet projects. Building something production-ready needs a harness.
  • Review is about risk. Review means identifying the risks and looking at what really matters, not reading every line.
  • LLM as a judge is worth having, with limits: it helps to write tests and to ease the review, but it doesn’t replace a human decision on risk.
  • Open question: are loops really a thing? With a good harness (docs, tests, rules), the agent should get it right quickly, so the loop should stay small.

Pitfalls

Section titled “Pitfalls”
  • Humanizing AI doesn’t work. Trying to reproduce human behaviour with persona agents (“the architect”, “the PM”…) does not work — don’t copy human processes onto agents.

Tools & tips

Section titled “Tools & tips”

MCP

Section titled “MCP”
  • Java MCP (javadocs.dev): gives the agent the Javadoc and sources of your dependencies, so it stops unzipping JARs. Endpoint: https://www.javadocs.dev/mcp

Skills

Section titled “Skills”
  • /grill-me and /grill-me-with-doc from Matt Pocock: the agent interviews you until the plan is clear. Setup guide

Commands

Section titled “Commands”
  • /goal: let the agent loop until a goal is reached.

Tools

Section titled “Tools”

Spec-Driven Development frameworks

Section titled “Spec-Driven Development frameworks”

Pick one framework and use it. Above all, change it to fit what you want to do: take ideas from it, modify it and make it your own.

  • OpenSpec: change proposals on top of living specs. Designed for brownfield and changes to existing code. openspec.dev
  • GitHub Spec Kit: step-by-step flow (constitution, specify, plan, tasks, implement). github.com/github/spec-kit
  • Kiro (AWS): spec-centric IDE, a VS Code fork by Amazon. kiro.dev
  • BMAD: a team of agent personas (analyst, PM, architect, dev). Produces a lot of .md. docs.bmad-method.org
  • AI Unified Process: built on use cases, aimed at enterprise, brownfield, long-lived systems. unifiedprocess.ai
  • Tessl: a skills registry for sharing versioned specs and skills. tessl.io
  • Matt Pocock’s skills: pragmatic, hand-crafted skills (grill-me, to-prd, wayfinder…). github.com/mattpocock/skills
  • Superpowers: brainstorming, multi-agent and adversarial review. github.com/obra/superpowers
  • GSD (Get Shit Done): a phase loop with each phase run in fresh-context subagents. github.com/open-gsd/gsd-core

Talks to share

Section titled “Talks to share”

If you watch only a few recordings, watch these — in this order.

Go further

Section titled “Go further”