Irrlicht.io is an open source tool for observing AI coding agents. But in this talk, the product is only the artifact. The real story is architectural: what happens when software is no longer built by developers, but by semi-autonomous agents?

What started as a small utility quickly became a living experiment in Software Architecture 3.0: Claude Code, Codex, Pi, OpenCode, and Gas Town forming a fast, fragile, surprisingly capable coding factory. Not a metaphorical factory, but an actual production system with throughput, bottlenecks, failure modes, coordination problems, governance needs, and architectural debt.

The talk explores the architectural questions that emerge when implementation speed explodes. Which responsibilities can be delegated to agents? Which ones must stay with humans? Where do we need boundaries, protocols, quality gates, observability, decision records, architectural fitness functions, and deliberate friction?

Some agents became reliable contributors. Others produced confident nonsense at impressive speed. But the core architectural tension remained the same: when does trust increase system throughput, and when does control preserve system integrity?

We will look at Irrlicht not as a product demo, but as a case study in designing socio-technical agent systems. The human role shifts from implementer to architect of constraints: defining intent, shaping feedback loops, designing review paths, managing blast radius, and deciding where autonomy is useful and where it becomes dangerous.

There is also a less glamorous side to high-speed architecture. What happens when the machine can continue producing work long after the human system is saturated? How do we design agentic workflows that are not only technically scalable, but cognitively and socially sustainable?

In the end, Irrlicht is the artifact that fell out of the experiment. The deeper topic is a new form of open source architecture: small autonomous software factories that can create astonishing output, but only when surrounded by strong architectural guardrails.

The central thesis: The next architectural superpower is not producing more code faster. It is designing systems of humans and agents that remain understandable, steerable, observable, and survivable.