Track 1

From NPCs to LLMs: Using game dev logic to build efficient code agents 

Caius Eugene

Current LLM agents are drowning in "token bloat." Shoveling entire repositories into a prompt doesn’t just burn budgets, it destroys reasoning. But what if we stopped treating code as a text block and started treating it as a navigable world?

Join me as I demonstrate how I’ve applied high-performance patterns from AAA Game Development to solve the context crisis. I’ll show you how I've replaced bloated prompts with Graph Traversals to prune data, Behavior Trees for reactive NPC-style decision-making, and Level of Detail (LOD) logic to balance abstraction with fidelity.

Through a live, interactive "code-world" simulation, we’ll explore how these game-engine tactics transform agents from passive readers into agile Pathfinders. Come learn how to build leaner, smarter autonomous systems that master repository-scale engineering without losing their mind.

What agents do when you’re not looking: Lessons from building and watching agent teams

Matias Vizcaino

Coding agents have evolved fast: from chat assistants, to copilots, to increasingly autonomous systems that can plan, delegate and act. That shift is exciting, but it also creates a new engineering problem: what are these systems actually doing when we are not watching closely? In this talk, I’ll trace that evolution, starting with war stories from the industry that make the risks feel real and familiar. Then I’ll share what I’ve learned from building and observing my own agent teams in Claude: the funny moments, the unexpected interactions, and the patterns that only become visible when you look at traces and behaviour instead of demos. I’ll close with the emerging practitioner playbook for this space: how teams are starting to observe, evaluate and govern agents so we can use them with more confidence in real-world delivery.

The augmented angineer: Evolving sensible defaults through Human-Centric AI 

Elena Guidi

Engineering excellence at Thoughtworks has always been defined by our Sensible Defaults,the collective wisdom that reduces decision fatigue and keeps us focused on value. Today, AI is redefining that baseline. While we are still actively defining what this new baseline looks like, the true shift isn't in the automation itself, but in why human practitioners remain more critical than ever.

In this session, I explore the personal experience of the "augmented engineer." We’ll move beyond the hype of agents to discuss AI as a disciplined tool for our daily craft. Drawing from real-world workflows, I will share how to integrate AI into your personal "defaults" with concrete examples, such as using it to scaffold complex legacy refactoring, or bridge niche tech stack gaps without sacrificing technical integrity. This is about elevating our human potential, ensuring that while the tools change, the curiosity, judgment, and responsibility remain firmly in our hands.

The Clarity-First Method: Vibe-coding production-ready AI Products without debt 

Maria Jose Lalama

Product Managers are becoming Product Builders. Vibe-coding and shipping AI-generated products fast is no longer an experiment, it's a core skill. PMs who can validate ideas quickly and reduce feedback loops become force multipliers for engineering teams. But the engineering concern is valid: Will this ship with technical debt?. Join me to answer the question with Vivian, a women's health app shipped in 48 hours using AI orchestration with structured product thinking. The result: zero technical debt, dedicated security sprints, and real users validating faster than ever. Built following the Clarity-First Method, that separates hastily-built prototypes from production-ready products.

Learn about Specification as Architecture to define constraints before AI builds, through research, PRD, system design, Master Prompts and Prompt Stories. Context-Aware Consistency to maintain quality across builds through token governance, security sprints, and human-in-the-loop decisions. And Strategic Reversions, reverting as discipline. Learn how clarity produces code engineers respect.

Evals are the new tests: Deploying an AI defect analysis system with confidence

Fabian Nonnenmacher

Moving an AI defect analysis tool from a "cool prototype" to a reliable production system takes more than just tweaking prompts. It requires a serious commitment to Evaluations (Eval). We’ll walk you through the real-world deployment of our system, treating evaluation as the equivalent of traditional software testing for AI applications.

Using our application as a live case study, we’ll dive into how we evaluated critical features like summarization, data extraction, and duplicate detection. We’ll demonstrate how an evaluation-driven development cycle turns the application's performance into a measurable, repeatable process rather than a guessing game. Finally, we’ll bridge the gap between the lab and the real world by showing how to combine offline benchmarks with live feedback. Our goal is to prove that running evaluations isn't a daunting academic exercise - It’s a necessity that brings the discipline of standard engineering to the world of AI.

How AWS thinks about spec-driven development

Matt Laver from AWS

With the growing adoption of coding agents and spec-driven development, how do you separate the hype from sustainable engineering practices? Early data suggests AI-assisted coding improves delivery speed but can compromise quality — spec-driven development aims to solve that tension by providing structure. In this fireside chat, AWS Senior Specialist Solutions Architect Matt Laver shares an inside look at how Kiro enables spec-driven development to help teams ship reliable software faster. We'll go beyond the tool itself to distill general principles for adopting SDD well — from shifting effort upstream into requirements and design, to maintaining delivery stability as teams scale their use of AI. Expect a candid discussion featuring genuine experience from real-world engagements in a still-developing discipline.

Track 2

LLM Wiki: Your mind, structured by AI

Maria Martinez Miralles

In the current AI landscape, systems treat every query as a blank slate, forcing them to rediscover information from scratch. The LLM Wiki pattern disrupts this by turning AI into a persistent knowledge architect. By transforming markdown tools like Obsidian into an active development environment, the AI treats a wiki as living source code—actively building and maintaining an evolving ecosystem of your data.

This shift replaces static, siloed documents with composite knowledge. As you add information, the AI automatically handles the operational "dirty work": summarizing, labeling, cross-referencing, and resolving contradictions. Consequently, your knowledge base becomes richer and more valuable over time. By delegating these tasks to the AI, humans are freed from information management and empowered to focus entirely on high-level exploration, strategic synthesis, and decision-making.

Breaking the sound barrier: Scaling ETL migrations to 300x with AI and engineering

Carolina Allende and Aakash Jain

Enterprises have been running critical data pipelines on aging platforms for decades and migrating them to GCP is genuinely hard.In this talk, we will share what actually happened when we combined engineering expertise with AI-assisted code translation on a large-scale migration. We built a human-in-the-loop workflow — not because it sounded good, but because full automation kept breaking things that mattered. The result: The first iteration we achieved 60x faster migration the goal was to ran 200–300× faster. The tradeoffs were real too. We’ll cover: ∙ What made the legacy platform so difficult to migrate ∙ How we designed the AI-assisted workflow and where we drew the human line ∙ What failed, what we fixed, and what we’d do differently ∙ What teams planning similar work should watch out for No vendor pitch. No cherry-picked metrics. Just delivery experience.

From Airspace to Apron: Closing the flight gate data quality gap 

Ruchika Rawat and Pemu Mungai from OAG

Ground-level flight data is one of aviation's most persistent data challenges — and one of its most valuable opportunities. Gate assignments, terminal information, and accurate arrival and departure records underpin the automated workflows that travel operators depend on, yet reliable, real-time gate data remains difficult to procure at scale.
This talk explores how Thoughtworks, in partnership with OAG — a global leader in travel data — developed an innovative methodology to infer gate assignments in near real-time. By combining live aircraft telemetry, airport apron geospatial data, and historical flight records, we built a scalable approach to closing critical data quality gaps without relying solely on carrier-originated feeds.
We will walk through the technical architecture, the aviation-specific challenges encountered along the way, and the strategies developed to overcome them — from positional data ambiguity to multi-level terminal inference.

Using AI to make legacy systems legible

Prachi Tyagi

Legacy systems are easy to dismiss as old technology, but in many organizations they are still the systems the business depends on every day. They carry years of decisions, exceptions, workarounds, integrations, and operational history. The difficult part is that this complexity is rarely visible in one place, which makes even small changes feel slower and riskier than they should.

This talk looks at AI through that legacy lens: not as a replacement for engineers, and not simply as a way to generate more code, but as a practical aid for making complex systems easier to navigate and safer to work with.

The session will offer a grounded way to think about where AI helps, where human judgment remains essential, and how teams can use both to evolve systems they cannot afford to break.

AI-First Software Maintenance - AI-First Software Delivery 

Vitória Regina Ragazzon Toebe and Amy Pellegrini

Most conversations about AI in software engineering focus on shipping features faster. But in a fast-moving environment, the real bottleneck often lives on the other side: the time between a reported issue and a confident resolution.
We'll share how a maintenance team and a delivery team — working on a core service with multiple consumers, spread across internal and external teams, investigating issues across multiple data sources — operate as one connected AI-first system. With workflows designed around AI strengths: triage, data correlation, supporting cross-team communication, and handling context cleanly between maintenance and delivery.
You'll leave with concrete patterns for AI-first incident management, a model for maintenance-delivery collaboration that scales, and an honest view of where AI helps — and where it still gets in the way.

Building sovereignty: Where your AI runs matters

Jens Peveling 

European organizations and governments racing to leverage AI and large-scale data analytics face an uncomfortable truth: the platforms they run their critical services on, train AI models on, the pipelines they feed sensitive data into, and the infrastructure governing their data residency are controlled by hyperscalers — architecturally unfit for true sovereignty. This talk is based on first-hand experience building a sovereign cloud from bare metal for the European energy sector. I present core concepts for building a platform designed for running critical workloads and data-intensive operations under full jurisdictional control — no hyperscaler dependencies, no shared-responsibility illusions. I examine why sovereign infrastructure is becoming a strategic necessity, not a niche concern: how data gravity, AI model training locality, and regulatory frameworks force a rethinking of where and how we build cloud platforms. The talk covers real trade-offs, real architecture decisions, and lessons learned the hard way.