What Occurred At AI-fokus 2026

AI-fokus, a spin off of Jfokus, is a convention focusing
on the evolution of agentic AI and its sensible implementation in software program improvement.

This 12 months’s second version coated a broad transition within the tech business; transferring past fundamental AI chatbots and preliminary prototypes to constructing, securing, and main sturdy, production-ready agentic techniques.

The core themes mentioned have been :

The evolution of Software program Engineering and Spec-Pushed improvement
A serious theme was the shifting id of the software program developer. As AI takes over the mechanical act of typing code, the human position is pivoting in the direction of defining issues, designing structure, and guaranteeing high quality.

Context Engineering
As a result of LLMs have finite short-term reminiscences, managing their context window has develop into a vital engineering self-discipline. Within the convention it was argued that context ought to be handled with the identical rigor as conventional code, requiring its personal lifecycle of technology, analysis, distribution, and commentary.

System Structure and the Mannequin Context Protocol (MCP)
Presenters explored the mechanics of integrating instruments into AI to make brokers autonomous. MCP was closely featured as the brand new customary for giving AI safe, standardized entry to exterior instruments, file techniques, and APIs. Presenters demonstrated tips on how to construct MCP purchasers in functions and tips on how to use MCP servers to attach LLMs to native knowledge with out polluting the core context window.

Safety, sandboxing, and infrastructure
As brokers achieve the power to execute code and modify environments, safety has develop into a paramount concern. Talks highlighted catastrophic failures, comparable to brokers by chance deleting manufacturing databases or buying unauthorized gadgets (like PlayStations) when given an excessive amount of unchecked entry.

To mitigate this, engineers are constructing safe infrastructure management planes. Options like Kubernetes Agent Sandbox enable brokers to execute generated code inside ephemeral, remoted environments (like gVisor) which might be destroyed instantly after use, neutralizing the blast radius of rogue code.

AI Testing, high quality, and observability
As a result of LLMs are non-deterministic and output probabilistic solutions, conventional unit testing is inadequate. Talks elaborated on what to do about it.

Management and Organizational Transformation
Lastly, audio system addressed the human component of integrating AI throughout giant enterprises.

Whereas many displays provided worthwhile insights, the next three talks stood out on AI’s affect on code high quality:

1. “Past the Vibe, How far can we push It?” by Roger Johansson

This speak stands out for taking AI’s uncooked coding capabilities to absolutely the excessive. Johansson got down to see if he might construct a completely purposeful, ECMAScript-compliant JavaScript engine totally by means of AI, with out writing a single line of code by hand.

As a substitute of straightforward prototypes, he efficiently navigated complicated architectural challenges by unleashing autonomous brokers in “Ralph loops”—scripts that repeatedly fed bug experiences and failing checks again to the AI in a single day.

Finally, his AI brokers have been in a position to parse and go over 93,500 checks from the official ECMAScript specification, culminating in an engine able to working a full NodeJS Specific internet utility domestically.

2. “How you can be a terrific chief to your Agentic groups” by Björn Wismén

Wismén’s presentation is an enchanting have a look at the way forward for organizational construction. His total firm—together with his management board and improvement squads—consists solely of himself and AI brokers.

What makes this speak stand out is his strategy to managing these fashions; slightly than simply giving them technical prompts, he packages them with distinct personalities and “tender abilities”. As an example, his C-suite brokers (CFO, CTO, CMO, CPO) are given particular ranges of “psychological security,” creativity, and stubbornness.

He demonstrated how he actually leaves the room whereas his AI board debates the monetary and technical deserves of taking over a brand new shopper undertaking, trusting them to reach at a sound enterprise determination autonomously.

3. “Agentic AI Coding Practices for Velocity with High quality” by Adam Tornhill

Amidst the hype of speedy AI code technology, Tornhill offered a significant, data-driven actuality verify concerning what he calls the “phantasm of prompt software program”.

This speak stands out as a result of it highlights the hidden risks of AI assistants. Tornhill shared analysis displaying that naive AI adoption can result in a 41% improve in software program defects and an enormous spike in AI-induced code complexity, which in the end makes initiatives slower to take care of after just some months.

To fight this “self-harm loop” the place AI generates code it can’t reliably keep later, Tornhill launched a extremely sensible answer: utilizing a MCP server to behave as an automatic gatekeeper that strictly enforces excessive code well being metrics earlier than accepting any AI-generated code.

Whereas these have been probably the most fascinating talks, there have been many different nice periods. So, take your time and delve into the convention by going by means of its YouTube playlist which incorporates all recorded periods.