The Winchester Mystery House Effect

Smart devices, convoluted code: Why mobile IoT modernisation using AI agents still needs human experts

Outdated mobile software architectures often resemble the Winchester House, which was extended bit by bit over the years by different people, according to their individual tastes and preferences. The result was a huge, absurd complex with 160 rooms and bizarre structural flaws.

The same is often true of old legacy applications. Due to constantly shifting responsibilities, tight budgets, intense time pressure, individual preferences and poor quality assurance, there is ultimately a lack of proper documentation. This is what makes it so difficult to maintain outdated software.

In many companies, the issues mentioned above hinder the further development of legacy systems. Maintenance ties up considerable resources. At the same time, technological dependency prevents the integration of modern IoT scenarios.

The use of autonomous AI agents and the adesso adSCAILE framework now opens up a viable path to transforming even complex and business-critical mobile applications at the interface with physical devices in a future-proof and cost-effective manner.

However, this does not mean that human involvement becomes obsolete. Quite the contrary: the so-called ‘human-in-the-loop’ plays a central and extremely important role.

They are becoming the digital chief architect.

I discussed and reflected on exactly what this means and how the use of agents affects the day-to-day work of a development team with three of my esteemed colleagues – Boris Beck, Chris Gretzki and Jens Schließer – from the development team. All of them have years of experience in the fields of software architecture, requirements, coding and testing.

Another topic that occupied us during the discussion was the question of what currently constitutes ‘marketing buzzwords’ and where AI brings real benefits. Particularly with regard to complex, business-critical applications that cannot simply be replaced with a bit of ‘vibe coding’.

For me, as a Senior Business Developer, this was a very important discussion, as I am involved in client meetings on a daily basis and, in my role, aim to provide advice – transparently and honestly.

I’d like to share these insights with you.

How autonomous AI agents work in the IoT environment

Unlike traditional, assistant-based code generators, agent-based systems can work proactively and contextually, given the right guidance. Autonomous AI agents analyse the entire software architecture of a mobile application. One could also say that analysing and documenting existing code is one of AI’s forte. In the IoT environment, which is characterised by error-free interaction with different firmware versions and a high degree of offline capability, this is a fundamental process.

A further benefit of AI lies in the ability to automatically analyse, digitally document and architecturally verify undocumented hardware interfaces and manufacturer-specific firmware dependencies. This ensures that the mobile application operates error-free and in compliance with regulations, even under unstable offline conditions or in the event of future firmware updates to the physical devices.

Strategic advantages through adSCAILE and specialised programming agents

With the adSCAILE framework (adesso Smart Cycle for AI-enhanced Lightweight Software Engineering), adesso provides a model- and platform-independent, end-to-end process.

This approach fundamentally changes the project economics in brownfield projects: instead of months-long analysis loops, adesso relies on ultra-short development and turnaround cycles, as well as smaller, highly efficient teams. Thanks to the automated generation of variants, both test coverage and quality within the existing architecture increase noticeably.

It is interesting to note that software development and the construction industry can be compared so closely — these analogies make technical concepts easy to understand even for non-programmers.

A ‘brownfield’ is a developed site — that is, a site with existing factories or buildings that need to be converted, refurbished or integrated. In software development, the term refers to legacy systems that are not developed from scratch, but are specifically modernised, expanded and improved.

For product managers and digital strategists, this results in three key advantages:

• Safeguarding the data flow
  Proprietary communication logic is systematically identified and extracted. This ensures that existing IoT data streams are preserved and can be transferred in a controlled manner to modern platforms and cloud infrastructures.

• Risk minimisation through automation
  The AI-supported comprehensive analysis of the existing source code reduces unforeseen implementation costs, highlights potential sources of problems and also provides complete documentation. This makes modernisation more predictable, both technically and economically. In practice, the use of AI results in significant savings in effort and time of between 30 and 50 per cent.

• Focus on value creation
  AI agents now handle repetitive tasks such as code generation, documentation and bug fixing. This frees up capacity for project teams, which they can then focus specifically on developing new business models, digital services and differentiating features.

Here at adesso, we use a range of AI tools, and are currently keen to utilise specialised programming agents such as Claude Code from Anthropic. Integrated directly into the development environment, the tool analyses the codebase, identifies dependencies on older network libraries and processes files across multiple levels.

It all sounds very promising – almost as if the agents were doing all the work on their own. The questions that inevitably arise are: are human developers now obsolete, and what justification is there for development costs under these circumstances?

These are legitimate and honest questions. Before the in-depth discussion with my colleagues, I too found it difficult to understand exactly how the use of AI actually affects our business.

What I learnt in my conversation with Jens, Chris and Boris.

• Why AI lacks an eye for the blueprint
  A key shortcoming of AI agents concerns semantics: the solutions generated by AI are rarely easy to read or maintainable in the long term at the initial stage. It simply lacks the strategic foresight to know what needs to be preserved in the code in the long term and in which direction an architecture should be allowed to develop. Without human input, AI would do exactly what Sarah Winchester did: it builds functional rooms but neglects the overall architectural plan. The more you ‘let the agents off the leash’, the more code ends up having to be validated manually.

• AI as a sparring partner
  Today, an experienced software architect tends to use the agent for a data-driven second opinion and plans major adjustments to classes or modules in collaboration with the AI. AI agents are also extremely valuable as sparring partners for working out requirements, often uncovering missing information during discussions.

• Working with AI agents shifts the focus of work within the software development lifecycle
  Away from manual programming and towards a highly detailed and precise requirements phase. The main focus is on creating very precise specifications. This means the design and planning phase takes on great importance. Work is carried out iteratively, step by step. Planning therefore takes longer, whilst coding takes less time.

• Those who code should not test
  This also applies to agents. After all, it is only logical that if an AI has a different understanding and is working in the wrong direction, it will test exactly what it has understood during testing. And that, as we all know, is not always correct. Here, too, it is human experts who check whether the tests have been written correctly and, consequently, whether the result can be correct.
• Definition:
  Agent-based development versus Vibe Coding

These two terms are often confused. In Vibe Coding, a broadly defined objective dominates the result, although nobody really knows how the AI arrived at that result. This is different in agentic application development, where the planning phase is significantly expanded, thereby creating the fundamental context for the overall solution.

My conclusions following the discussion with my colleagues are as follows: AI may be fast in certain areas and offer some added value, but without very specific guidance and control over the results, even AI generates many errors.

And no matter how fast the AI is, the internal coordination effort with specialist departments within the client organisation, the specification of functions and human biorhythms still only allow for a certain pace.

Conclusion

The modernisation of mobile IoT applications is a strategic prerequisite for securing and expanding the market position of connected products. Agent-based software development significantly reduces the risks and effort involved in this transformation. Companies overcome technological bottlenecks, increase the adaptability of their system landscapes and lay the foundations for continuously developing smart products and positioning them in the market in a future-proof manner.

However, without the professional and dedicated work of human subject matter experts, even an AI project is no guarantee of success.