AI-Powered Software Creation: Trends and Strategies for 2025
The emergence of Large Language Models (LLMs) is revolutionising software development at an unprecedented pace. Their generative capabilities in producing code and, on a broader scale, entire new products, are reminiscent of the Industrial Revolution. Novel methods of creating digital assets are emerging, potentially eliminating certain jobs while creating others. Combine this with disruption similar to the birth of smartphones, along with the technological innovation introduced by cloud computing, and you have a glimpse of our current era.
With all these activities and the continuous improvements in AI research, these are three expectations for 2025 and my related playbook.
Entrepreneur acceleration will require different engineering decisions
LLMs’ code/product generation capabilities will take centre stage in MVPs and prototypes. With improved time to market and less capital needed to start due to smaller engineering teams, LLMs will make entrepreneurial ideas a reality at an unprecedented pace. While it might be easy to believe we are in the “no engineer” era, I think 2025 is still the “low code” era. With the GenAI space still fluid, featuring a spread of tools and approaches, it’s critical to have a strategy to navigate decisions. My playbook for GenAI product creation for early-stage businesses would be built around two core topics:
- GenAI code technology integration capabilities with human-in-the-loop. The ability to continuously hand off code/product between LLMs and humans will be key for true acceleration. Is the chosen tool a black box/walled garden, or does it natively support a human-LLM hybrid workflow?
- Team expertise in “driving” decisions (instead of passively accepting LLMs decisions). While I believe smaller teams will be needed to achieve MVP stage, LLMs still hallucinate (yet less and less) and can go into an infinite bug creation loop (fix one problem and create a new one). Experienced engineers are the ones who can break the cycle and assess in real-time if the generated code is fit for the problem, making the acceleration real. We are also looking at engineers who are comfortable with the emerging style of development: conversational development with AI is here to stay.
AI platform: protocols and data contracts to the rescue
While more opportunities around AI emerge, the technology stack around LLMs is in its infancy. For example, Agentic AI is what everyone is betting on, but creating agents and agentic workflows is still complex. A set of standards is emerging, simplifying the ‘how’ and fuelling additional technical innovation at the platform level. The ModelContextProtocol introduced by Anthropic feels like a step in the right direction. In 2025, I would consider open ecosystem integration when assessing frameworks and tools. Similarly, data contracts and types will become more important than ever. For certain problem spaces, LLMs can already take over from humans. Your data contracts are critical to enable the human-AI hybrid development flow.
Future of Software Engineering: known techniques, in a new context
LLMs will shine in transforming new ideas into products, with smaller teams achieving great MVP stage, but working with mature products will still require more human expertise than LLMs. This will be particularly true for legacy systems, where internal quality can be low, making product evolution a game of whack-a-mole. LLMs can’t understand whole systems (including distributed patterns), and even less can resonate effectively on a chain of poor decisions. My headcount budgeting would be “traditional”, taking the 20% out-of-the-box productivity boost coming from AI-assisted development as a bonus. I would invest in hardening the system boundaries to facilitate incremental swaps towards the AI future.
- Techniques such as TDD and pair programming will see increased adoption, with an LLM twist
- Driving expected behaviours from an LLM is most effective when combined with a TDD-style approach, involving defining the end state, iterating on prompts, refactoring, and creating tests. Pair programming with LLMs will likely become standard practice, extending beyond mere autocompletion to encompass a continuous dialogue with an AI agent throughout the feature development lifecycle. As these changes unfold, architecture’s role in designing future system growth trajectories will become increasingly central to development discussions.
I see 2025 as a transitional year, with a hybrid human-AI assisted way to create and evolve products. I’m an innovation optimist, so I expect software engineering to be completely transformed positively in the next couple of years. But which job will not be transformed by AI?
