The Future of Software Development with (Generative) AI- Part 1

Author(s): Amir Piltan

Originally published on Towards AI.

Introduction

Over the past decade, AI techniques — particularly deep learning — have been increasingly applied to create AI-Assisted Development (AIAD) tools designed to support developers across various stages of the Software Development Life Cycle (SDLC). These tools initially focused on enhancing productivity and quality through intelligent code completion, bug detection, and basic automation.

However, with the rise of Generative AI, this domain has gained unprecedented momentum. In fact, software development has become one of the most impactful and heavily invested use cases of Generative AI to date. As Generative AI continues to evolve in the coming years, it is poised to fundamentally reshape the entire SDLC. From requirements gathering and code generation to testing, deployment, and maintenance. A June 2023 study estimates the AI code generation market at USD 30.9 billion, projected to reach USD 169.2 billion by 2032. Gartner predicts that “by 2028, 90% of enterprise software engineers will use AI code assistants, up from less than 14% in early 2024.”

With over five years of hands-on experience in designing and building AI-Assisted Development (AIAD) products, I would like to share my perspective on the evolution of AIAD — its early foundations, current trajectory, and where I believe it’s headed. This series will explore the product landscape, its growing impact, and a forward-looking vision of how AI will transform software development, redefine the role of developers and Integrated Development Environments (IDEs), and reshape the broader software industry. The focus will remain on product innovation, developer experience, strategic implications, and future possibilities, rather than diving into the deep technical mechanics behind these products.

The Evolution of AI in Software Development

The birth of AI-Assisted Development Tools

In-context code suggestion and completion have been among the first generation of the AI-Assisted Development (AIAD) capabilities. A notable early example was Visual Studio IntelliSense, introduced in 1996 and continuously improved over time, followed by tools like Eclipse Code Completion, which was phased out in 2019.

Although a bit out of fashion in recent years, in-context code suggestion and completion tools have evolved using enhanced AI techniques. For instance, the latest version of JetBrains Full Line Code Completion goes beyond simple suggestions by recommending complete lines of code. Unlike Generative AI-based AIAD, this tool’s models runs locally without sending code over the internet, which can be faster, cheaper and more compliant. There have been low code equivalent of Code suggestion/completion such as Mendix Logic Recommender that we built in Mendix low code Integrated Development Environment (IDE), Studio Pro. These tools typically combine language semantics, heuristic rules, and repository/project context to provide in-context code suggestions. However, they lack the comprehensive ‘code generation’ or ‘conversational’ capabilities to interact with developers as generative AI models do.

Despite significant efforts and enthusiasm, the adoption and productivity impact of the first generation of AIAD tools have remained relatively limited.

Generative AI and the Second Phase of AIAD: AI Development Bots

With the advent of Generative AI in 2023, AI-Assisted Development (AIAD) has entered a second phase, evolving into way more powerful development assistant tools with broader and deeper capabilities — what I call as AI Development Bots. Notable examples include Github Copilot, Gitlab Duo, Amazon Q Developer, IBM watsonx Code Assistant, Qodo, Tabnine and low code versions of them such as Mendix Artificial Intelligence Assistance (Maia) and PoweApps Copilot. These tools are typically delivered as conversational chat UIs integrated into IDE sidebars, enabling developers to interact with them using natural language — that’s why I call them AI Development Bots.

The common capabilities of these Large Language Model (LLM)-powered chatbots include answering coding-related questions, explaining code, troubleshooting, suggesting next steps, and, most notably, generating code snippets based on natural language descriptions. Their impact and adoption have surpassed those of the previous generation. For example, GitHub Copilot, the world’s most widely adopted AI Development Bot, has reached 1.3 million paid subscribers (a 30% increase since June 2023) across 50,000 organizations, with a recent 30% quarter-over-quarter growth. GitHub reports that 88% of developers “feel more productive” 60% “feel more fulfilled” and developers accept 30% of its code suggestions. This marks a significant leap from the first the generation of the AIAD tools.

However, while AI Development Bots are more intelligent, powerful, and context-aware than their predecessors, they still act as assistants to human developers. I can remember that once they were introduced, many questioned why do we need developers if AI can build software?! Despite the buzz, these bots are not autonomous and can hardly complete an end-to-end development task or user story as human developers can. They often generate generic code snippets and lack the reasoning abilities of humans. At their best, these bots guide developers and produce helpful code snippets — which is undeniably impressive — but code is not a working application. This is what the next generation AIAD tools might do!

AI Agentic Frameworks and the Third Phase of AIAD: AI Development Agents

In the next phase, which began in 2024, AI Development Bots are evolving into AI Development Agents, gaining more capability and autonomy. These agents are becoming more than just tools or bots — they are closer to AI agents, though not yet comparable to human developers. That’s why I name them as AI Development Agents. They have the potential to take direct actions to complete simple development tasks or user stories based on natural language requirements.

A typical workflow could be like an AI Development Agent analyzing a requirement in natural language, gathering project data and context, breaking the story/task into smaller sub-tasks, executing those actions by writing code, committing the code for human developer code review, iterating based on feedback, and commit the code again. They go beyond the AI development bots stream by not only generating code, but also configuring environments, installing packages, setting up the project structure and creating entire modules based on predefined templates. Additionally, they generate test cases aligned with the application’s code and specifications, as well as test scripts for unit, integration, and end-to-end testing. These agents are powered by two key advancements: improved Large Language Models (LLMs) and the rise of AI agent frameworks. Andrew Ng, a prominent AI expert, predicts that “AI agent workflows will drive massive AI progress this year (2024) — perhaps even more than the next generation of foundation models.” This is where software innovators like GitHub, emerging GenAI start-ups, and Generative AI researchers are currently focusing on.

One early example is Devin AI, introduced by Cognition Labs in March 2024. According to Cognition Labs, “Devin can plan and execute complex engineering tasks requiring thousands of decisions, recall relevant context at every step, learn over time, and fix mistakes.” I personally have no experience with Devin to evaluate to what extent it is autonomous and capable of doing development tasks. There are also open-source alternatives of it such as OpenDevin, Devika, and SWE-Agent.

Another early development is Github Copilot Workspace, launched in April 2024. What I really like about this tool is how it keeps human developers in control and engaged throughout the process — a challenging yet vital aspect of designing AI-agentic workflows.

Multi-Agent Collaboration, a new frontier in AI development agents: An emerging design and architecture pattern that is taking development agents to the next level is multi-agent collaboration. This approach breaks down complex tasks, like writing software, into smaller subtasks executed by multiple agents. Each agent may utilise one or more specialised Large Language Models (LLMs). Frameworks like AutoGen, Crew AI, and LangGraph are at the forefront of enabling this collaborative paradigm. Using its multi-agent conversation framework, Auogen enables multi-agent creation and collaboration where agents that are capable of conversing with each other through the exchange of messages to jointly finish a development task. Agents can be enabled by different LLMs, act with different levels of autonomy and have configurable human (developer) involvement.

The recently introduced GitHub Copilot Agent mode (April 2025) is as a clear example of the agentic development. Github claims that it is an autonomous and agentic real-time, synchronous collaborator that performs multistep coding tasks based on natural-language prompts. It analyzes your codebase to grasp the full context, build multistep solutions, run tests, use external tools for specialised tasks. The key agentic feature is its capability to iterate on its own code, recognizing errors, and fixing them automatically through an agentic loop. Cursor Agent Mode and Augment Code are other popular examples of the AI Development Agents. They independently explore (e.g., read documentation, browse the web, etc.), plan, edit files and executes coding tasks. Windsurf Wave 2 and 3 operate in a real agentic way, they search the web for answers and parse any supplied URL. As the developer keep talking to the agent, it will start recording memories, learn rules and use them moving forward.

Initial experiments suggest that multi-agent systems outperform single agents in coding tasks, as demonstrated by results on the widely recognised HumanEval coding benchmark. Andrew Ng cites that by incorporating an iterative agent workflow, GPT-3.5 (zero shot) performance on code writing jumped from 48.1% to up to 95.1%.

Latency, costs and performance, challenges and opportunities: Latency and costs, driven by the multiplied API calls required for multi-agent collaboration, remain significant challenges in using these tools in production. However, the cost per million tokens for LLMs is rapidly decreasing, and latency continues to drop. The trillion-dollar race to develop AI-specific hardware is driving this progress. In addition, I believe the current capabilities of Generative AI and LLMs, while remarkable, are still limited to helping developers become more productive rather than replacing them. In my opinion, by no means we are near the state where Generative AI can independently create software and replace human developers.

The underlying reason is that despite the significant progress in AI, LLMs are still limited in what they can do. As Yann LeCun, VP & Chief AI Scientist at Meta, has pointed out, the LLMs fundamentally lack the ability to truly understand context, apply common sense reasoning, grasp complex real-world dynamics, or innovate beyond their training data. Instead, they often hallucinate responses, generating content that lacks factual grounding or deeper comprehension.

However, each new model is expected to be smarter in remarkable ways, and we may even see the emergence of entirely new types of Generative AI models or systems beyond LLMs.How will this ongoing evolution impact software development? Let’s explore the next phase.

The Near Future of AI in Software Development: From AI Development Agents to AI Co-Developers

So far, we’ve explored the state of AI in software development, both in the past and the present. The next phases, however, are more speculative — my predictions about the future. I envision the next generation of Generative AIAD evolving into AI Co-developers, which already feels almost like a real developer teammate.

Much like a human developer, these AI co-developers could pick up or be assigned user stories during a sprint. Being virtual, they could work 24/7, leaving human developers to review and iterate their work before deployment. I’ve intentionally distinguished the evolution into AI agents and AI co-developers to emphasize the gradual nature of this progress. Initially, we may see robotic-like AI agents with limited capabilities, which will incrementally evolve into more sophisticated, human-like AI co-developers.

The key differences between AI agents and AI co-developers can be explained by what we earlier quoted from LeCun: the ability to truly understand context, apply common sense reasoning, and grasp complex real-world dynamics. To achieve this, emerging startups like Poolside and Magic are working on training specialized LLMs from the ground up for software development tasks. For example, Magic’s LTM-2-mini can handle up to 100 million tokens in a single inference. This equates to approximately 10 million lines of code, enabling it to analyze an entire code repository, documentation, and libraries in context when suggesting or generating code. This contextual depth is essential for tackling the complexities of real-world software development. The second pillar that differentiates AI co-developers from AI agents is their advanced user experience and AIAD-native IDEs to accommodate effective collaboration between human and AI developers. I will explain this in the next part of this series.

One important note about the prefix “Co-” in AI co-developers is to highlight their limited autonomy at this stage. They operate under strict human supervision, with human developers maintaining full control. However, as capabilities evolve, this dynamic might shift. Let’s explore what could come next.

The Ultimate Future of AI in Software Development: Virtual AI Development Teams

As AI technologies continue to advance, AI co-developers will gradually become more autonomous, capable of grasping complex real-world dynamics. I argue that this progression could lead to the next phase: Specialized AI Development Teams. These would be virtual teams composed of AI developers covering various roles, much like real development teams of human developers, which include backend developers, frontend developers, database engineers, data engineers, product owners, and more.

Arriving at this phase would require significant advancements in Generative AI, specifically AI systems that can learn, remember, reason, and plan, referred to as “Objective-Driven AI” by LeCun VP & Chief AI Scientist at Meta. He explains “such AI as systems that develop rich internal world models capturing how things work, interact and change. They would be capable of simulating outcomes, reasoning about the future, and making informed decisions to achieve specific objectives”, exactly like real human development teams.

The underlying AI models in this phase would likely be trained or fine-tuned for specific software development tasks or roles. Unlike current approaches, where LLMs are trained on massive codebases, future training data might focus on recordings (textual or video) of real human development teams in action. These datasets would capture not only code but the entire development process, including collaboration, decision-making, and project management activities.

When combined with frameworks like AutoGen — or possibly new ones — these specialized AI systems could enable the realization of nearly autonomous virtual AI development teams.

A Paradigm Shift: This is a huge paradigm shift from the previous phases. Specialized AI developers won’t just assist human developers; they’ll take the lead, building entire applications end-to-end — not merely writing code snippets or delivering single user stories. This marks the transition from an AI-Assisted Development (AIAD) paradigm, where human developers are in the lead and AI provides assistance, to a Human-Assisted Development (HAD) paradigm, where virtual AI developers take the lead, and human developers play a supporting role.

I speculate that shift will likely transform the roles of human developers and the structure of software development teams. A new human role will emerge — what I call the Development Orchestrator. This role will focus on:

1. Clarifying and shaping requirements: Working closely with business stakeholders or customers to refine requirements.
2. Supervising and orchestrating AI Development Teams: Guiding AI developers and managing the development process.
3. Validating code and applications: Ensuring the generated code and applications meet quality, functionality, and business needs, in a new era of human-machine collaboration.

Conclusion

In the this part of this series, I have shared the evolution of AIAD — its early foundations, current trajectory, and where I believe it’s headed. We explored how AI is radically transforming the way software is built, the evolving role of developers, and its broader impact. In the next part, I’ll present a bold vision for how integrated development environments (IDEs) will evolve with AI, and what that means for developer roles and productivity. I’ll also discuss product and development collaboration in the age of AI and speculate on emerging business models in the software industry. Stay tuned!

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI