How do we keep control over AI-generated code? The first AI Engineer Europe in London offered no simple answer. To make the positions easier to grasp, I have grouped them into three approaches: Slow Down, Speed and Open Source. Reality is messier than that, naturally, but as Hayden White put it: "The beginning of all understanding is classification."
Three of the headline speakers had Austrian roots: Peter Steinberger (OpenAI / OpenClaw Foundation), Mario Zechner and Armin Ronacher. Together with Cristina Poncela Cubeiro (the latter three now at Earendil), they were among the defining voices of the conference.
The assessments and conclusions in this article are my own, based on the talks and the conversations I had on-site. I only cover sessions I attended in person (and believe I understood). I have simplified things in places, so please flag any errors!
Event Details
Event
AI Engineer Europe 2026 – Europe's first official AI Engineer conference, organised by Swyx (Shawn Wang) and the AI Engineer team
Dates
8–10 April 2026 — Day 1: Workshops + Expo, Days 2–3: Keynotes + Breakouts + Afterparties
Location
Queen Elizabeth II Centre, Broad Sanctuary, Westminster, London SW1P 3EE – directly opposite Westminster Abbey
Attendees & Format
1,000+ AI Engineers from around the world, 100+ speakers, 11 technical tracks, 23 hands-on workshops (Day 1)
Table of Contents
The Conference at a Glance
Click loads YouTube (Privacy)
Day 1 (9 April) – Agents, Security and Economics. Malte Ubl on falling code costs, Raia Hadsell on DeepMind beyond language, Peter Steinberger on OpenClaw's security explosion, and Matt Pocock on software fundamentals in the AI age.
Click loads YouTube (Privacy)
Day 2 (10 April) – Code Quality and Open Source. David Soria Parra on the future of MCP, Mario Zechner and Armin Ronacher & Cristina Poncela Cubeiro on agent-legible architectures, Sarah Chieng on fast models – and Tuomas Artman (Linear) on taste as a competitive advantage.
Approach 1: Slow Down
This group shares a clear conviction: AI speed without human control produces not innovation but uncontrollable junk. Their remedy: clean software design and clear lines of responsibility in the code.
Matt Pocock – Software Fundamentals are More Important Than Ever
Code is not cheap. Bad code is more expensive today than ever because it prevents AI from working efficiently.
Matt Pocock took apart the widespread "Specs-to-Code" approach: the practice of simply throwing specifications (detailed requirement descriptions) at an AI and adopting the output blindly. The result, he argues, is uncontrollable "software entropy" (the creeping rise of disorder and complexity in a codebase).
His counter-proposal:
- Deep Modules (a concept from John Ousterhout): modules that hold a lot of logic internally but expose a simple interface. Like a car: a complex engine, but all you need is a steering wheel and pedals. The human designs the architecture; the AI implements the logic inside the module. Testing happens only at the interface, which makes it ideal for TDD as a guardrail.
- "Grill me" prompts: he makes the AI cross-examine him until they reach a shared understanding of the system (the design concept), rather than churning out code blind.
- TDD: tests first, code second. This lets him steer the AI in small, verifiable steps.
Mario Zechner – Self-Control and the Fight Against AI Junk
- A human safeguard: developers feel a natural resistance when code turns messy, and sooner or later they stop to tidy it up.
- AI lacks this instinct: its only goal is code that runs and passes the tests, by whatever means.
- Bad role models: agents fill the gaps in a task description with patterns from the internet, which are largely made up of our own technical debt.
- The result: because agents never pause, they pile up in days the chaos that would take humans months, until the codebase is so cluttered that eventually even the agents can no longer fix it.
The Austrian developer Mario Zechner (best known as the creator of libGDX) built the minimalist coding agent Pi because commercial tools like Claude Code take context control away from developers. Pi ships deliberately with only four tools: read, write, edit, bash.
Three central points from the talk:
-
Automated agents are destroying Open Source: agents flood open-source projects with useless pull requests. Zechner built a junk filter that closes PRs automatically and demands human confirmation. When it all gets too much, he takes an "OSS Vacation" and shuts the issue tracker.
-
No pain, no quality: unlike humans, agents don't learn organically from their mistakes. The intellectual friction of programming is essential to genuinely understanding a system.
-
"Slow the fuck down": critical code still has to be read by a human. Slowing down isn't a step backwards; it's quality assurance.
Armin Ronacher & Cristina Poncela Cubeiro – "The Friction is Your Judgement"
Armin Ronacher, creator of Flask and, after a decade at Sentry, now founder of Earendil, delivered one of the conference's sharpest analyses alongside Cristina Poncela Cubeiro. Mario Zechner has also been part of Earendil since April 2026.
The efficiency trap: AI tools are addictive. You never know whether the next prompt will deliver the feature or fill the system with junk. Speed creates the illusion of productivity. In reality, there's no time left to think: code reviews get waved through because agents produce more code than humans can ever verify.
Working code ≠ good code: a human developer feels uneasy writing messy workarounds. An agent doesn't. It writes code that "runs somehow", for example failing silently with default values on error instead of aborting cleanly. The result: corrupt data nobody notices and technical debt that grows fast.
Libraries yes, complex products no: AI agents can write well-defined software libraries with clear boundaries and simple interfaces. On complex products, where the interface, permissions and billing all interlock, the AI lacks the big picture. It behaves sensibly locally but makes the system as a whole steadily more chaotic.
The "Agent-Legible Codebase": code that agents can read without going wrong:
- Each function does exactly one thing, and its name makes that clear
- What the code says is what actually happens, with no hidden side effects (e.g. React Server Actions or complex ORMs that obscure the code's intent from the agent)
- Automated checks (linters) prevent, for instance, empty error handlers that swallow problems
- Error messages describe the specific problem ("Database connection failed") rather than just "An error has occurred"
Deliberate friction: an AI agent can change hundreds of files in seconds. That's great for routine work, but dangerous when it rebuilds a database structure or alters access rights along the way. Ronacher draws an analogy with physics: without friction, there's no steering. So they have built tools (extensions for the Pi agent) that bring critical operations to a hard stop. The human gets a clear summary: "I want to drop table X and change permissions for Y, shall I proceed?" Nothing continues without explicit approval. That friction is the moment when human experience and judgement come into play.
Ronacher and Poncela Cubeiro's conclusion: AI is a brilliant tool for narrowly defined tasks, such as reproducing bugs. But architectural decisions belong in human hands.
Read more on webconsulting.at
Making TYPO3 AI-ready: How clear structures deliver better agent results – Agent-Legible Codebase in practice: How we made a TYPO3 codebase readable for AI agents.
Approach 2: Speed
This group wants quality too, but instead of hitting the brakes, they want to build infrastructure so that quality emerges systemically even at high speed.
Malte Ubl – Mass Software Production
Software production has become so cheap that we can now build all the software that was previously uneconomical to develop. This will massively increase the demand for developers – not decrease it.
Malte Ubl (CTO of Vercel, the cloud platform behind Next.js for hosting, serverless and edge computing) supplied the economic perspective:
-
Falling infrastructure costs: AI inference costs are dropping fast, driven by competition between Google, OpenAI and other providers. Ubl's point: once infrastructure is cheap, the money is made by those who build good products, not those who run the servers.
-
Agents as primary users: at Vercel, agents already account for over 60% of page views. In future, infrastructure will need to be optimised natively for agents (APIs/CLIs) rather than for human UIs.
-
Agents as a new application layer: bespoke automation is economically viable for the first time. It's no longer just large enterprises; small teams can now build custom agents too.
Swyx – Tiny Teams and AI Automation
Swyx (Shawn Wang), the conference organiser and founder of AI Engineer, showed live how his nine-person team generates over 9 million dollars, handling work that used to take entire departments.
-
The end of "Yak Shaving": before you can get to the actual task, you first have to fix ten other things; that's Yak Shaving. Agents take over this tedious groundwork.
-
Replacing SaaS: he uses AI to replace complex SaaS products (such as a CMS) outright with AI-managed code.
-
Everyday usefulness: agents for everyday tasks, including web research to track down a real lobster in London. No joke.
Read more on webconsulting.at
AI-Native Companies: How developers work with AI agents – How tiny teams integrate AI agents into their daily routines using the PDAA workflow.
Keeping AI Costs Under Control: The practical guide to strategic budget planning – What happens when code becomes almost free – and where the hidden costs lie.
Approach 3: Open Source
Peter Steinberger – State of the Claw
Three capabilities that are manageable on their own, but together form a complete attack chain: reading personal data (the loot), processing untrusted websites (the attack surface) and sending emails (the escape route). Steinberger covered this in his talk.
Peter Steinberger, founder of PSPDFKit, at OpenAI since February 2026 and founder of the OpenClaw Foundation, gave one of the most talked-about sessions with "State of the Claw". OpenClaw is built on Zechner's Pi, has grown explosively and has racked up over 1,142 Security Advisories.
For comparison (from Steinberger's talk):
| Project | Timeframe | CVEs / Advisories | Rate |
|---|---|---|---|
| Django | 19 years | 94 | ~8/year |
| curl | 8 years | ~600 Reports | ~100/year |
| Linux Kernel | Since 2023 (CNA change) | ~8–9/day | Sharply increased |
| OpenClaw | 68 days | 1,142 total | 16.8/day |
- Independence: to stop large corporations from seizing control, the OpenClaw Foundation is being set up as a "neutral Switzerland", along the lines of the Linux Foundation.
Mario Zechner – Pi as an Open-Source Alternative
Zechner embodies both approaches at once: Slow Down and Open Source. His agent Pi is living proof that, alongside proprietary products like Claude Code, Cursor and GitHub Copilot, there is a radically different path: a lean, open-source tool that leaves developers in full control of their workflow.
Proprietary coding agents decide behind closed doors which context they send, which background actions they run and how they handle your data. An open-source agent like Pi makes these decisions transparent and auditable, which is essential for security-critical or data-sensitive projects.
For Pi, this isn't a compromise but a deliberate design principle: fewer features, more control. No magic, no hidden API calls, no telemetry. Exactly four tools, and full responsibility stays with the human. The fact that Pi now forms the technical foundation of OpenClaw (Steinberger's project) shows that minimalism and scalability aren't mutually exclusive.
Swyx in Conversation with Steinberger – Local Control and "Taste"
In an open AMA format, Swyx and Steinberger discussed the philosophical foundations:
-
Control over your own data: local models (AI running directly on your own machine rather than in the cloud) make it possible to sidestep the data silos of the big tech companies. Rather than waiting for official APIs, an agent can simply work through web interfaces.
-
"Taste" as a differentiator: as AI automates the sheer production of code, value shifts to human "taste", the ability to tell whether a design or a piece of code "reeks of AI" or genuinely has a soul.
Read more on webconsulting.at
Model Context Protocol: 30 Questions and Answers – MCP is the open standard through which agents like Pi and OpenClaw communicate with external tools.
Agent Skills: 30 Questions and Answers – The open standard for reusable AI agent capabilities in Claude Code, Cursor, and VS Code.
Synthesis: Three Approaches, One Goal
Every speaker wants quality; there was consensus on that. The differences lie not in the goal but in the weighting. The three approaches aren't mutually exclusive, and in practice most teams will probably combine elements of all three.
Slow Down
Quality through human control
Slow down, keep intellectual control, build in deliberate hurdles. AI-generated code needs human verification.
Speed
Quality through better systems
Build infrastructure so agents work cleanly from the start. API-first, validation, automated feedback loops.
Open Source
Quality through transparency
Build tools openly and locally. Anyone who can see the code can verify it. Proprietary black boxes are the opposite.
Bonus: Tuomas Artman (Linear) – Why "Taste" is the New Engineering
The fireside chat between Tuomas Artman (CTO and co-founder of Linear) and Gergely Orosz (The Pragmatic Engineer) was my personal highlight of Day 2. You rarely hear someone argue so clearly from their own practice. Artman brought a perspective that none of the three approaches covers on its own; instead it ties them all together.
In a world where anyone can generate code with AI, "taste", the instinct for good design, the right abstractions and consistent opinionation, becomes the decisive competitive advantage.
The most important statements:
-
Zero Bug Policy: at Linear, bugs take top priority; nobody starts on new features until they're fixed. Artman's logic: every bug gets fixed eventually, so why not now? The result: customers report a bug and find it fixed the very next day.
-
Opinionated software (software with a clear point of view): Linear is deliberately not flexible enough for every workflow. There is one good way to do things, and the system guides you towards it. This cuts decision fatigue and increases speed.
-
Linear Agent: Linear builds AI straight into the product, not as a chatbot but as an AI PM that handles triage, backlog grooming and issue creation. On top of that come deep integrations with coding agents via MCP, so you can launch a local agent session straight from an issue.
-
Craftsmanship over metrics: Linear leans on good design and gut instinct rather than endless metrics and optimisation loops.
-
Five-day hiring: anyone applying to Linear comes in for five days and builds something real. No whiteboard interviews, no LeetCode, just real work on the real product. It filters precisely for the kind of person Linear wants: people with a love of detail and craftsmanship.
Personal note: alongside Mario Zechner, the conversation with Artman impressed me the most. Everything at Linear is built with a love of design you can feel straight away, from the product to the hiring process. Honestly, I wouldn't say no to a five-day trial at Linear.
Modern code reviews with structural diffing that vastly reduce the number of changed lines in many cases.
Review, comment, check previews, get notifications on failed CI check.
Statischer Tweet-Inhalt – keine Cookies, kein Tracking
What Else Moves the Speakers
What the conference speakers posted in the days surrounding the event:
Peter Steinberger – Anthropic Locks Out Open Source
Both me and @davemorin tried to talk sense into Anthropic, best we managed was delaying this for a week.
Funny how timings match up, first they copy some popular features into their closed harness, then they lock out open source.
Statischer Tweet-Inhalt – keine Cookies, kein Tracking
Mario Zechner – "I've sold out"
mariozechner.at/posts/2026-04-08-ive-sold-out/
Statischer Tweet-Inhalt – keine Cookies, kein Tracking
Armin Ronacher & Cristina Poncela Cubeiro – Earendil Explodes
Statischer Tweet-Inhalt – keine Cookies, kein Tracking
Malte Ubl – just-bash Keeps Delivering
And the best thing: Because we have a unified Fluid Compute stack across Sandbox, Builds, and Functions these wins.
Statischer Tweet-Inhalt – keine Cookies, kein Tracking
My Conclusion
My impression after two days in London: the question is shifting from whether AI agents will transform software development to how we keep control while they do.
The sharpest analyses came from the people who build coding agents themselves: Steinberger (OpenAI / OpenClaw Foundation), Zechner, Ronacher and Poncela Cubeiro (Earendil). Their message: technology needs guardrails, Open Source needs protection from spam, and code needs humans who understand it.
My five personal takeaways:
-
Open-source alternatives are not a luxury but a necessity. The major providers (Anthropic, OpenAI, Google) are presumably pursuing a winner-takes-all strategy. Build your workflow on a proprietary tool and you become dependent on someone else's decisions. Steinberger's experience with the Anthropic block makes the point: access can be cut off overnight. Projects like Pi and OpenClaw are therefore not niche products; they are a strategic safeguard against platform lock-in.
-
How much faster are you really? The honest answer is probably a factor of 2 to 3, not the 10x some people claim. Zechner sums it up perfectly: in practice, being ten times faster often just means producing ten times as much junk. You only become genuinely more productive if quality control keeps pace, and that takes discipline, not speed. (Note: this phrasing probably traces back to a tweet I can no longer find; I'm afraid I conflated it with the talk.)
-
Google's pricing strategy could be a game-changer. If Ubl is right and AI inference costs keep falling, value creation is likely to shift ever further towards the application layer. For agencies and developers, that would be an opportunity.
-
"Taste" could become the most valuable currency. The more code machines generate, the more important human judgement about what is good, as opposed to what merely "works", is likely to become.
-
Friction is not a bug but a feature. Ronacher and Poncela Cubeiro are right: when an agent wants to run a database migration or a permissions change, it has to stop and wait for human approval rather than simply pressing on.
My Addition: Software Creation is not Software Operations
The conference revolved almost entirely around the creation of code. In practice, though, I ask my clients other questions first:
- Who handles the infrastructure updates, and how fast does the team react if a dependency is compromised? (Case in point: the Axios Trojan of March 2026, a North Korean attack on an npm package with 70 million weekly downloads that smuggled a Remote Access Trojan onto thousands of systems within hours.)
- Who owns backup and disaster recovery?
- How quickly does a critical bug get fixed in production?
- And, above all: what does the data strategy look like? Is correct, up-to-date data for the whole company available at the push of a button?
Without clean, validated live data, the whole AI effort is pointless. My advice: organise, clean and make the data accessible first, then we can talk about agents. And yes, agents can help with the data cleaning itself too: finding duplicates, standardising formats, spotting missing fields. But someone has to set the strategy and check the results.
AI agents can only ever be as good as the data they're given. If your data isn't under control (inconsistent master data, stale exports, manual workarounds), even the best coding agent won't work miracles. Data quality comes before agent quality.
The first AI Engineer Europe set the bar high. Technical depth, strategic foresight and a community that brings both enthusiasm and critical distance: you rarely find all of that in one place. If the second edition keeps up this standard, London will become a fixture in the calendar.
Day 1: Full Programme
Click loads YouTube (Privacy)
Clickable chapter markers (each timestamp links straight into the video):
- 00:13:10 – Opening Remarks by Phil Hawksworth
- 00:21:26 – Lia McBride (AI Engineer): 900% community growth and British AI infrastructure investments
- 00:24:25 – Malte Ubl (Vercel): Agents as a new application layer, APIs must become "AI first"
- 00:42:39 – Raia Hadsell (Google DeepMind): Gemini Embeddings 2, AI cyclone forecasting, Project Genie 3
- 01:07:08 – Ryan Lopopolo (OpenAI): "Code is free" – Systems thinking and delegation for parallel AI agents
- 01:25:48 – Peter Steinberger (OpenAI / OpenClaw Foundation) 🇦🇹: "State of the Claw" – OpenClaw's explosive growth and AI-generated security bounty flood
- 01:45:12 – Break: Morning Coffee
- 02:28:13 – Swyx with Peter Steinberger (OpenAI / OpenClaw Foundation) 🇦🇹: Open Source, "Token Maxing" and "Taste" as the ultimate engineering moat
- 02:55:01 – Vincent Koc (Comet ML): "Dark Factories" – 60+ parallel AI agents for nocturnal codebase refactoring
- 03:14:07 – Radek Sienkiewicz (VelvetShark): Handing over personal life to OpenClaw – via Obsidian, email, and background tasks
- 03:34:12 – Sally Ann O'Malley (Red Hat): Secure agent deployments with Podman, Docker, and K8s – isolation and state recovery
- 03:57:05 – Nick Taylor (Pomerium): Securing OpenClaw with Identity-Aware Proxy and live-coding an MCP server from Discord
- 04:14:35 – Break: Lunch
- 05:41:51 – Onur Solmaz (OpenClaw): ACP for standardised agent interactions and disposable agents on K8s
- 06:02:17 – Merve Noyan (Hugging Face): HF ecosystem for local coding agents and model training via Hub Skills
- 06:22:36 – Fryderyk Wiatrowski (Viktor): "Viktor" – Slack-native AI employee with context across thousands of integrated tools
- 06:42:09 – Break: Afternoon
- 07:42:39 – Gergely Orosz (The Pragmatic Engineer) with Swyx: "Token Maxing" – Big Tech engineers wasting AI inference to inflate productivity metrics
- 08:09:26 – Kitze (Sizzy): Modern productivity apps roasted – and an OS where AI generates the UI on demand
- 08:29:42 – Matt Pocock (AI Hero): Why DDD and TDD are more important than ever against AI-generated "slop"
- 08:48:31 – Sunil Pai (Cloudflare): "Code Mode" – LLMs executing JavaScript in V8 isolates, bypassing slow JSON tool calls
- 09:07:04 – Closing Remarks by Phil Hawksworth
Day 2: Full Programme
Click loads YouTube (Privacy)
Clickable chapter markers:
- 00:10:40 – Tejas Kumar opens Day 2
- 00:15:44 – Omar Sanseviero (Google DeepMind): Gemma 4 on-device capabilities and E2B architecture
- 00:31:00 – David Soria Parra (Anthropic): The future of MCP and programmatic tool calls
- 00:49:44 – Ido Salomon (MCP Apps): AgentCraft and the visual orchestration of multi-agent coding swarms
- 01:01:05 – Mario Zechner (Earendil / Pi) 🇦🇹: The Pi agent and the dangers of AI-generated technical debt
- 01:19:33 – Armin Ronacher & Cristina Poncela Cubeiro (Earendil) 🇦🇹: Agent-legible codebases and conscious friction
- 01:38:12 – Benjamin Dunphy: AI Engineer World's Fair announcement
- 01:44:14 – Break: Morning Coffee
- 02:26:10 – David Gomes (Cursor): Replacing 15,000 lines of code with Markdown Skills and Git Worktrees
- 02:46:17 – Matthias Luebken (TAVON): Embedding OpenClaw and Pi in multichannel production environments
- 03:08:39 – Sarah Chieng (Cerebras): Adapting developer habits for ultra-fast models like Codex Spark (1,200 TPS)
- 03:27:11 – Lawrence Jones (Incident.io): AI for evaluation, debugging, and management of complex AI systems
- 03:45:47 – Luke Alvoeiro (Factory): Architecture for long-running, multi-day agent missions
- 04:04:47 – Break: Lunch
- 05:41:46 – Ben Burtenshaw (Hugging Face): Coding agents for AI systems engineering and CUDA kernel development
- 06:00:33 – Michael Richman (Cmd+Ctrl): Curing FOMAT with mobile command-and-control
- 06:17:29 – Liam Hampton (Microsoft): Orchestrating local, background, and cloud agents simultaneously in VS Code
- 06:35:28 – Break: Afternoon
- 07:41:28 – Tuomas Artman (Linear) with Gergely Orosz: Fireside Chat on Linear's design philosophy and Zero Bug Policy
- 08:10:48 – Jacob Lauritzen (Legora): Vertical AI – why complex agents need permanent UI artefacts instead of chat
- 08:25:11 – Peter Gostev (Arena AI): The "Bullshit Benchmark" and what top models on LMSYS Arena still cannot do
- 08:45:32 – Swyx: Automating a 9-million-dollar conference business with AI agents for non-coding tasks
- 08:59:02 – Closing Remarks by Tejas Kumar
Read more on webconsulting.at
Code at a Crossroads: 7 Insights from the AI Engineer Summit 2025 – Our report from the predecessor event in the USA: War on Slop, Skills Architecture, and Agent-Ready Codebases.
From Coder to Orchestrator: What the Anthropic Report Means for Teams – How the Software Development Lifecycle is changing due to multi-agent systems.
TYPO3 Extension Security: What We Can Learn from Cloudflare's EmDash – Capability Manifests as a security model for agent permissions.
Impressions from London
London welcomed us with brilliant sunshine. The Queen Elizabeth II Centre sits in the heart of Westminster, right next to the Abbey, a stone's throw from Big Ben, with the London Eye and the Houses of Parliament in view. A conference venue could hardly be better placed.
Arrival & City
Arrival at Paddington Station – the Victorian roof structure from 1854 welcomes travellers
Architecture detail: Isambard Kingdom Brunel's masterpiece of steel and glass
Light installation in Paddington – modern lighting design meets Victorian architecture
Big Ben in the spring sun – just a few minutes' walk from the conference centre
London Eye and River Thames – view from Westminster Bridge in the sunshine
Contrasts: London's modern skyline behind the historic Vauxhall Bridge
Houses of Parliament – direct neighbour of the Queen Elizabeth II Centre
Winston Churchill watches over Parliament Square – a few steps away from the conference
St James's Park – perfect lunch break between the sessions
Plane tree avenue in St James's Park – London's green lung next to Westminster
Parliament Square – vibrant life at the political centre of London
Typically London: Houses of Parliament under a cloud-covered sky
London Eye close up – coffee break at Gail's Bakery
View from the hotel towards Westminster – Big Ben through the glass atrium
The Cenotaph in Whitehall – Britain's central war memorial, MCMXIX
10 Downing Street – heavily guarded and yet photogenic
Conference, Keynotes & Talks
The programme at a glance – welcome display at the QEII Centre
Behind the scenes – breakout room with professional AV setup
Packed house – the main hall during the keynotes (not everyone got in)
Concentrated audience – the community between the sessions
Peter Steinberger (OpenAI / OpenClaw Foundation) on the main stage – State of the Claw
Keynote on the main stage – the sponsor wall shows the Who's Who of the AI industry
Vercel title slide on the main stage – shortly before Malte Ubl's keynote
Raia Hadsell (Google DeepMind): From Atari to Robotics – the path to AGI via games and simulation
Neuroscience meets AI: Jennifer Aniston Cells – selective neuron activation
Peter Steinberger showing the security advisory explosion: OpenClaw 16.8 advisories per day
David Soria Parra (Anthropic) – the inventor of the Model Context Protocol (MCP) on the main stage
Coding Agents Track – one of the most attended tracks of the conference
The sponsor wall in all its glory – the biggest names in the AI industry under one roof
The core message: 2026 is all about connectivity – the best agents use everything
Another perspective of the imposing main stage
Armin Ronacher and Cristina Poncela Cubeiro (Earendil) – Agent-legible codebases and conscious friction
Slides & Insights
Sonar LLM Leaderboard – 50+ models ranked by code quality and security
The Top 5: Gemini and Claude Opus dominate in code quality and pass rate
Sarah Chieng (Cerebras): Fast Models Need Slow Developers – provocation as a programme
GitHub numbers: From 1 billion to 14 billion commits – AI agents drive the growth
Stanford 120K study: Clean code amplifies AI gains – the Slow Down faction feels validated
BullshitBench: Claude dominates when it comes to pushing back against false instructions
The three-way battle: Anthropic vs. OpenAI vs. Google – Claude leads in bullshit detection
What's the gap? – where even the best models still systematically fail
Fireside Chat – Gergely Orosz (The Pragmatic Engineer) in conversation with Tuomas Artman (Linear)
Jacob Lauritzen (Legora) – AI agents need genuine user interfaces, not just chat
Soundtrack to the Conference
Three songs that, for me, capture the mood of this conference.
Peter Gabriel – Solsbury Hill
Click loads YouTube (Privacy)
Peter Gabriel describes his own experience on the hill of the title in Somerset, England, a moment that gave him the courage to leave Genesis and strike out into the unknown. A good fit for a conference where so much was called into question.
Lou Reed – There Is No Time
Click loads YouTube (Privacy)
Real music: you can't beat two guitars, bass and drums. Lou Reed, raw and direct. A reminder that some things simply stay human, and are all the better for it.
Elton John – Tiny Dancer
Click loads YouTube (Privacy)
From England to California, 1970, and utterly relaxed. The perfect song for the flight home, as it all sinks in and London disappears behind the clouds.