AI Is Changing Software Engineering. The Fundamentals Still Matter.

I was fortunate to join almost 70 people at the European Future of Software Engineering (FOSE) retreat, hosted by Thoughtworks. AI dominated every conversation at FOSE. Surprisingly, though, the biggest lesson wasn't about better models or autonomous agents. It was that the organisations moving fastest are investing in better boundaries.

Across the diverse experiences, there was also a theme of "slow is smooth and smooth is fast". Investing in the right primitives unlocks tremendous opportunity, but many are still figuring out which primitives they are and how best to invest in them.

This event brought together thoughtful people from across regions, organisations, and disciplines to explore the future of software engineering in the age of AI. Under the Chatham House Rule, I can’t share specifics without prior approval, but many attendees have already published their own personal reflections, and I highly recommend reading them (some examples: 1, 2, 3).

Turns out taking trains with fellow participants just means extending the event! One final view of the venue before traveling back home.

Here is my summary of the event.

Providing enabling constraints is as important as ever

Primitives are most impactful when they are right-sized for use, composition, and orchestration. When working with AI, chunking context to the right size, time and time again, proved to be a worthwhile investment. One session was explicitly dedicated to code modularity, but the theme was also present as an undertone in nearly all sessions due to its broad applicability. For example:

• Right-sizing the human teams enables fast decision-making and exploration

• Chunking the code base so that the LLM has efficient token use and more effective solutioning

• Scoping the ecosystem and access for agents enables safer and more powerful solutions

In one of the many hallway track moments, an attendee provided proof. They ran a test in code by refactoring a large single-file codebase into more modular functions and later files. The token usage and solution quality when implementing a new feature in that codebase improved substantially as the file sizes became smaller and more targeted.

We can't slow down consumption, but we can enable it safely and efficiently

A shift in who traditionally consumes software-focused tooling was a key theme. The question is no longer if non-engineers will build software; it is how much. And the engineering teams (and specifically the platform, SRE, and operational teams) need to figure out how to support this rather than remove it. Accessing easy-to-deploy PaaS and cloud computing environments couldn't constrain software engineers 10 years ago, but it put a similar pressure on organisations. They all of a sudden had to support the request for these powerful tools or realise the impact of shadow IT. That is happening again today in the shape of vibe-coded apps, and we once again have to figure out how to support it rather than pretend we can avoid it, as shadow IT is already happening.

I have long advocated for APIs first; UIs may be an enhancement, but are insufficient on their own. In the past, I would point at CI/CD workflows as an almost universal understanding of an important consumer that can't effectively consume web-based GUIs. The need to invest in backend services first, with consumer interfaces as enhancements, is becoming even more clear with the use of Agentic AI. Yes, the chat window is what helped OpenAI burst onto the scene, but even non-engineer users now depend on high-quality MCP access to their tools to maintain value, which is fundamentally shifting user and vendor behaviour and unlocking outcomes.

One of the attendees had a catchy phrase for this, but the best I can remember is "vibe to value", which I don't think is quite right. Either way, the intention is captured in that organisations need to figure out how to provide tools that provide safety and efficiency in ongoing iterations while being coherently consumed by software engineers, non-engineer humans, deterministic automation, and agentic automation.

The industry is still enjoying expansion mode, but maintenance mode is coming for us

The majority of the event felt like a reinforcement of the gold rush mentality, focusing on how to do more with AI. This even showed up in the session that asked us to think about day-2 and the world where "they build it, you run it", where the conversation started by focusing on AI SRE and self‑healing agents. But many cautioned that while there are very clear wins happening in the issue identification, context providing, and solution suggesting space, automated remediation is still something to make intentional decisions around. The lack of an "embodied mind" and human-style learning means that removing humans from the loop also removes organisational and product learning. This may make sense in some contexts, but may also hamper harness and product improvements in the long term.

While the gold rush is clearly underway, many voices in the room noted that winning required spending more time thinking about the growing maintenance burden that follows. Engineering leaders are looking at the stability and automatability of support in the old data centre world for inspiration and seeing that clear actions that are deterministic, scoped to specific scenarios, and tightly role-controlled are what made operations so much easier to automate and support at scale. So while the landscape is far more dynamic and expansive today, the goal is to create operational harnesses that can support any deployed software and its operational and infrastructure dependencies, thereby unlocking more efficient support. While the focus right now is heavily on FinOps and token economics, the learnings from this space are also being applied to expectations for technical resource support.

The fundamentals are becoming more important, not less

AI is changing almost every aspect of software engineering. It is changing who produces software, who consumes it, and how it is maintained. What isn't changing is the importance of good engineering fundamentals. Well-designed abstractions, clear interfaces, modular systems and operational discipline aren't relics of a pre-AI world. They're becoming the foundation that allows humans and AI to build software together safely and sustainably.

Just like the first version of this event spawned this European edition, this event should spawn many more conversations like this that span regions, domains, and roles. We must keep trying to spot the emerging trends that can make AI not just effective in a capitalist organisation, but supportive in human endeavours, and sustainable in our environment.