Open-weight models: What are they and when should you use them?

We've seen a huge amount of increase in interest around open-weight models in 2026. The reasons for this are multifaceted, ranging from the switch to per token billing, the evolving privacy landscape and improvements in open-weight model capabilities. To unpack open-weight models and explore why and when we might want to use them, host Ken Mugrage is joined by Thoughtworkers Caer Sanders and Andre Almar. They discuss everything from the precise definition of the term and how they compare to other kinds of models, to governance and hardware challenges.

We've been thinking a lot recently about the status of code in a world where it's increasingly written and read by AI. Could the spec, for instance, become the primary way we interface with software systems? And if it does, will that mean humans no longer need to look at code at all?  At present, the idea of not looking at our code seems fanciful, possibly dangerous, but even within Thoughtworks there are conflicting perspectives about whether this might change. To some, code will always remain the primary artifact; to others, its role in how we build software is about to undergo a significant transformation. On this episode of the Technology Podcast, Thoughtworks' Caer Sanders and Razin Memon join host Ken Mugrage for a debate about how the relationship between developers and code might — or might not — change in the final years of this decade. For Caer, despite the clear capabilities of AI, paying attention to code will always matter; for Razin, meanwhile, code's relevance is likely to decline as harnesses become more sophisticated and specification techniques evolve.  Whatever your view, listen for a frank and open discussion about an issue that will ultimately determine what the future of software engineering actually looks like. Listen to our episode on harness engineering from May 2026: https://www.thoughtworks.com/insights/podcasts/technology-podcasts/what-harness-engineering Read a recent blog post on thoughtworks.com arguing we still need to design code for humans: https://www.thoughtworks.com/insights/blog/programming-languages/should-still-design-code-humans

Harness engineering is still a new concept, but already we've noticed a challenge being consistently faced by technology leaders: how can a harness be used at scale across an organization to ensure consistency and controllability without undermining autonomy? To some extent it's a discussion that's reminiscent of platform engineering and, before that, DevOps, but, given the nature of AI agents, it's also unique to this particular moment. On this episode of the Technology Podcast, host Ken Mugrage is joined by guests Thomas Squeo (Chief Technology Officer and Head of Advisory for the Americas) and Matt Kamelman (Innovation Choreographer) to explore why it's so important to scale harnesses effectively and the steps and processes that will allow you to do just that. Thomas and Matt recently wrote an article outlining how organizations should think about what they call 'enterprise harness engineering'; here, they unpack their ideas and explore what's required of engineering leaders and their teams. Read Thomas' and Matt's article: https://www.thoughtworks.com/insights/articles/operating-system-enterprise-ai

Cloud was one of the main drivers of the early waves of AI adoption. However, as AI has become more and more embedded in systems and devices — in both consumer and enterprise contexts — it's becoming a bottleneck. This isn't just about costs (although yes, that issue is certainly surging up the agenda), it's also about how we optimize our architectures and improve device performance. This is why conversation is turning to hybrid AI: embracing a hybrid approach that combines proprietary cloud services with local or on-device AI can help organizations deliver better experiences for users, whether they're consumers, other businesses or internal teams. One company exploring this space is electronics giant Lenovo. In this episode of the Technology Podcast, host Prem Chandrasekaran is joined by Girish Hoogar, Lenovo's Global Head of Technology for Cloud and Software, to discuss why the company is embracing hybrid AI, how it's approaching implementation and what the implications are for other technologists. As industry attention turns to local AI, listen for a first-hand perspective on what the trend actually means for engineering teams and their organizations.

At the end of June, Thoughtworks and Martin Fowler convened an unconference-style event in Switzerland with a range of industry leaders. The aim was to reflect on the current challenges and learning around AI assisted and agentic software engineering and discuss the implications for the future — and, most importantly, what actions need to be taken today. To review the event and dive into some of the topics that surfaced, host Ken Mugrage is joined by Thoughtworks colleagues Kief Morris (author of Infrastructure as Code) and Andrew Harmel-Law (author of Facilitating Software Architecture). They highlight some of the key ideas and issues that emerged from the event's conversations, ranging from developer identity to harness engineering. Learn more about the European edition of the Future of Software Engineering Retreat: https://www.thoughtworks.com/about-us/events/the-future-of-software-development-europe-2026 Read Andrew's recent blog post that asks whether non-functional requirements could be a vital guardrail for AI-generated code: https://www.thoughtworks.com/insights/blog/architecture/non-functional-requirements-missing-guardrail-ai-generated-code

What is code? It might sound obvious, but if you scratch the surface it becomes more difficult to articulate precisely what we mean. AI is complicating the picture further and changing the relationship developers have with code: when large amounts of executable code can be generated from high-level descriptions, what does it even mean to write code?  On this episode of the Technology Podcast, host Alexey Boas is joined by Thoughworks Distinguished Engineer Unmesh Joshi to discuss what code actually is and what it means to write, test, review and maintain code today. Building on Unmesh's recent article 'What is Code?' for martinfowler.com, this discussion dives into one of the fundamental building blocks of software while also thinking through the implications for 2026's engineering challenges. Read Unmesh's article on martinfowler.com: https://martinfowler.com/articles/what-is-code.html

Database branching has, for a long time, been a troublesome piece in the modern developer workflow puzzle: a good idea in principle but in practice a slow and often expensive challenge. Get it right and you can accelerate productivity and remove bottlenecks; get it wrong and you're potentially creating all sorts of trouble for yourself, from privacy risks to additional complexity. However, things are changing. Thanks to the emergence of new platforms such as Neon, Supabase and Databricks Lakebase, branching a database can become as familiar to developers as managing code branches and multiple environments with, say, Git and Terraform.  On this episode of the Technology Podcast, host Ken Mugrage is joined by his Thoughtworks colleague Cam Casher and Databricks' Kevin Hartman to discuss the work Thoughtworks and Databricks have been doing together on Lakebase. They discuss the platform, their experience using it with Spotify's Backstage and the opportunities database branching can offer software engineering teams in an increasingly AI-assisted and agentic world. Read Cam and Kevin's recent series on using Databricks Lakebase with Backstage: https://www.thoughtworks.com/insights/blog/data-engineering/backstage-lakebase-databricks

Semantic diffusion, combined with the pace of technology change, makes talking about AI-adjacent practices and techniques incredibly diffficult. There are few better examples of this issue than the term 'spec-driven development'. Although it's not new — its coinage precedes our current AI moment — it has become ubiquitous over the last six months or so as software professionals attempt to develop a vocabulary for talking about how they're developing methods for working successfully with coding agents. On this episode of the Technology Podcast, Birgitta Böckeler is joined by Laura Tacho — Developer Experience at AWS — to discuss all things spec-driven development. From competing definitions to different interpretations, implementations and workflows, the discussion provides a frank and grounded look at one of the most discussed and debated terms in modern software engineering. Learn more about Laura's work by visiting her website: https://lauratacho.com/ Read Birgitta's article on spec-driven development on Martin Fowler's website: https://martinfowler.com/articles/exploring-gen-ai/sdd-3-tools.html Learn more about The Future of Software Development Retreat discussed on this episode and explore some of the key insights: https://www.thoughtworks.com/about-us/events/the-future-of-software-development

'Harness engineering' is one of the most significant terms to emerge in software engineering in 2026. Broadly referring to the work done to control unpredictable AI agents and coding assistants, its use signals growing attention on what needs to be done to make agents reliable and consistent enough for production software in the real-world. On this episode of the Technology Podcast, Birgitta Böckeler joins hosts Prem Chandrasekaran and Nate Schutta to explore what harness engineering actually is, how it should be done and why it should matter to software engineers working today. Having written a number of articles on harness engineering for martinfowler.com based on her experiences with AI-assistance, Birgitta is well-placed to explain the core concepts and implications. Taking in everything from the practices and ideas that pre-date and inform harness engineering to integrating harness engineering into existing workflows, listen for a conversation that will provide much needed clarity on what's an essential topic in the industry. Read Birgitta's article on harness engineering on martinfowler.com: https://martinfowler.com/articles/harness-engineering.html Watch Birgitta's video on harness engineering beyond skills on YouTube: https://www.youtube.com/watch?v=uLWOLmeHOSE 

Anthropic Mythos garnered significant attention when it was launched in mid-April 2026. Yet despite it apparently presenting an unprecedented threat to global software, you don't have to look to closely to see that this was an effective product launch as much as a story about the grave security risks of today's AI models.  But this isn't to say there aren't important implications for software developers, security professionals and other technologists. In this episode of the Technology Podcast, one of our new hosts Nate Schutta is joined by Chris Kramer to discuss Anthropic Mythos and Project Glasswing, unpacking what's hype and what really matters. A few links for this episode: Some more information about Project Glasswing: https://www.anthropic.com/glasswing A story about how a small Discord group briefly had access to Mythos: https://daringfireball.net/linked/2026/04/23/discord-group-has-claude-mythos-access How Mozilla used Mythos to discover Firefox bugs: https://www.wired.com/story/mozilla-used-anthropics-mythos-to-find-271-bugs-in-firefox/