When an AI agent fails in production, the model isn’t always the problem. Brittle context handling, weak tool contracts, missing evaluation loops, and guardrails added too late can undermine even the most capable models.

This session makes the case for harness engineering as a first-class delivery discipline, exploring the components and design patterns behind reliable agents.

Drawing on real-world examples, it shows why simply switching to a bigger model can mask deeper problems and where teams should focus their effort when building agents that need to perform beyond the demo.

AI agents are changing how technology teams work, taking on tasks that have traditionally been part of the learning journey for junior technologists. As productivity accelerates, the path to learning and expertise is changing too.

This panel explores the tension between AI-amplified performance and genuine expertise, the experiences that remain essential for developing critical thinking and judgement, and the evolving meaning of seniority in technology. The conversation will also examine how these changes could reshape access and opportunity, creating new pathways for some while potentially introducing new barriers for the next generation of tech talent.

In this session, we'll explore how AI can be used in software development processes, with a focus on spec-driven development. We'll discuss how this approach can help teams accelerate development and how to get started with its adoption.

LATAM Airlines reduced P85 Time to Market by 69%, from 275 days to 85 days, across more than 20 squads using an AI-native product development lifecycle. But the biggest impact didn’t come from generating code. It happened earlier, in Discovery and Definition. This session shares the results of a six-month study, including audited metrics and validated hypotheses, showing how AI helped improve decision artifacts and reduce Build time by 84%. It offers a practical, evidence-based look at scaling AI across software delivery with measurable business results.

As AI moves from chatbots to autonomous agents, the challenge is shifting from smarter models to the data they rely on. Agentic systems need more than searchable information. They need data that is trusted, contextual, traceable, governed and ready to support real-world actions. This session explores the data foundations behind dependable agentic AI, from architecture and governance to the operational practices that help move AI beyond promising demos and into resilient production systems.

A backlog of more than 3.8 million unclassified images, documents, and videos would have required nearly 100,000 hours of manual work, creating a significant operational burden at enterprise scale. This session explores how a production-grade machine learning workflow automated asset classification with approximately 97% accuracy, using a secure, resilient, and event-driven cloud architecture built to process millions of assets without disrupting existing services. It also shares how the solution became a reusable foundation for bringing machine learning into enterprise products and enabling future intelligent automation.

Grupo Guanabara began modernizing a complex ticket pricing system involving machine learning, real-time decisions, business rules, and specialized expertise. But the solution became the starting point for something bigger: building the technological foundations of a data-driven organization. This session explores how a modular architecture on AWS, a modern MLOps ecosystem, and shared AI capabilities helped move beyond a standalone machine learning initiative, creating a foundation for new AI products, intelligent agents, and human-in-the-loop decision-making.

What if you could test what works before committing to a major technology investment? Working with a Fortune 50 retailer on a $400M vendor income leakage opportunity, the team used small, AI-assisted experiments to test data readiness, contract capture and leakage scenarios before building anything. This session explores how that experiment-led approach turned a complex opportunity into a validated, prioritized roadmap in weeks rather than quarters, and offers a repeatable way to de-risk large-scale platform investments before writing production code.

GenAI is changing who builds software and what they expect from internal platforms. As functional applications can increasingly be created using natural language, the challenge shifts from building a prototype to getting it safely and rapidly into production. Drawing on real-world lessons from Thoughtworks and Meta, this session explores how platform engineering is evolving from reusable components toward agentic skills, orchestration and governance, enabling greater speed without compromising security or quality.

AI investment doesn’t automatically translate into business value.  Even well-funded initiatives can stall after rollout if people aren’t using AI confidently and competently in their daily work.

Organizations seeing results aren’t just deploying AI faster. They’re focusing on how adoption actually happens. This keynote explores the human side of AI and the connection between adoption, utilization and ROI, and how organizations can bridge the gap between AI investment and measurable business value.

AI adoption isn't the finish line.  Business value is.

Design systems were built for human designers and developers. Now AI agents are designing interfaces, writing code, reviewing pull requests and maintaining products, becoming a new kind of consumer of those systems. Drawing on experiments building an agentic design system, this session explores how design systems can evolve beyond component libraries into structured knowledge AI can understand and act on, why traditional documentation may no longer be enough, and the principles teams can use to make their design systems AI-ready.

AI is accelerating software development, but quality feedback can still be slow, complex and increasingly non-deterministic. Adding more automation or LLM-based evaluations can help, but it can also introduce another feedback loop to maintain. This session explores a practical path from probabilistic evaluations to deterministic checks embedded in the software development lifecycle. It looks at how teams can use evaluations to understand complex quality problems, identify what can be verified deterministically, and turn those insights into fast, repeatable and trustworthy checks that keep pace with AI-assisted and agentic software delivery.

AI can generate code in seconds, but building software you can trust is still the hard part. After working with AI-powered products, coding agents and context-aware platforms, the limits of vibe coding become clear as code generation accelerates and intuition alone stops being enough. Drawing on real-world successes, failures and lessons learned, this session explores the shift from prompt to context engineering, why manual review no longer scales, and how specifications, automated validation and deterministic engineering practices can help keep AI-assisted software reliable, maintainable and production-ready.

Modern observability assumes you can put an agent on the machine. But when customers lease GPU infrastructure, that assumption breaks: the hardware is yours to keep running, yet the contract says you can't log in. This session shares lessons from building observability for a GPU cloud provider running more than 18,000 GPUs. It explores signals across GPUs, hosts, network fabrics, power and liquid cooling: what remains observable without in-band access, what lower-fidelity alternatives cost, and which questions become impossible to answer.

Veja o AI/works em ação em uma demonstração prática focada na modernização de sistemas existentes. Utilizando uma aplicação Java do setor de saúde, esta sessão explora como o AI/works analisa o código existente para identificar a arquitetura do sistema, suas dependências e regras de negócio, transformando esse conhecimento em especificações estruturadas que podem apoiar os esforços de modernização.

Conheça a abordagem da Thoughtworks para o desenvolvimento com agentes de IA e descubra como o AI/works pode ajudar as equipes a compreender melhor sistemas complexos e acelerar sua modernização. Ao final, haverá espaço para perguntas e respostas.