MAD About Software Design - When AI Architects Argue

The rise of large language models and autonomous agents has turned some familiar software-design processes upside down. In this talk we’ll examine the idea that instead of a single AI assistant, you could have multiple agents debating architecture - challenging one another, surfacing assumptions, exposing trade-offs, and helping drive better decisions. Building on the ideas presented in “MAD About Software Design: When AI Debates”

We’ll cover three major areas:
1. Why this matters: Scaling software development is more than scaling the actual coding itself. We already look at areas of writing specifications, testing, coding and deploying with AI-driven agents. Software design is inherently a human task. But can we also scale that with AI?

2. How the “multi-agent debate” model works: Agents with different roles (e.g., “performance champion”, “security skeptic”, “maintenance guardian”) argue and negotiate design proposals. We’ll walk through what such a system can look like, its pitfalls, and how it differs from just “AI writes code for you”.

3. Lessons learned from real-world evaluation & implementation: Abstract ideas are fun, but it’s not enough for real-world impact. We’ll talk about implementation, experiments, failures, and learnings.

4. Practical guidance for your team: We’ll talk about practical implementation, and where this pattern and relevant tooling fits in the larger scheme of SDLC. What to look out for, and where the humans fit in all of this.

Beyond the Context Window: Building a "Mental Model" for Your AI Agents

We’ve all seen what happens when you give an LLM a single repo: it’s usually quite successful.
But what happens when you give it a distributed system with 200 microservices? It gets lost in the "muddy middle".
The bottleneck in AI-driven development isn't code generation; it’s architectural context.

This talk introduces the Application Architecture Hub - a living, AI-maintained knowledge base designed specifically for agent consumption.

We will explore how to move away from static, stale documentation and toward a system of automated "facets" (APIs, Domain Models, Events, and more) that allow AI agents to troubleshoot, design, and assess impact across your entire ecosystem.
We'll explore how this can benefit workflows beyond coding, across the organization.

Attendees will leave with a blueprint for building an ingestion pipeline that keeps their AI agents (and their humans) in sync with the "truth" of the codebase.

Should about 30min talk, can be more.
Assumes basic understanding of LLMs and coding agents, and development background.