Technology & Communications
Autonomous software development
Agentic AI is moving software development from automated code generation toward fully autonomous management of entire development cycles.
While traditional GenAI tools have already automated tasks like code suggestions, code generation and debugging, the next evolution - autonomous software development - will involve AI agents proactively analysing requirements, independently planning software architecture, executing comprehensive coding tasks, and autonomously conducting testing, debugging, and deployment activities with minimal human supervision.
These kinds of agentic AI systems continuously interact with development environments, APIs, and external resources, dynamically adapting their approaches based on real-time feedback, outcomes of earlier deployments, and changing project parameters. This autonomy accelerates software delivery cycles and improves resource allocation, allowing developers to focus primarily on high-level oversight and strategic planning instead of routine coding or maintenance tasks.
Copyright: who owns the code of the future?
The advancing autonomy of agentic AI in software development - compared to just using GenAI - is putting the traditional principles of copyright law to an even greater test.
At its core, the use of agentic AI raises three main copyright questions:
- Are the work products of an agentic AI eligible for IP protection?
- Does the agentic AI infringe third party IP rights while processing the task on its own e.g. while doing scraping processes?
- Does the agentic AI infringe third party IP while interacting with third parties e.g. interacting with third party licensors?
To enhance our understanding, this section focuses solely on the question of protectability.
Agentic AI marks a shift toward autonomous systems that execute entire processes with minimal, and in some cases no human input. This boosts efficiency but also endangers copyright protection by completely detaching people from the creative process loop (e.g. coding the software). In purely internal code-generation workflows, the absence of copyright protection may be tolerable. However, for companies licensing code to third parties, the protectability of fully or partially AI-generated code is essential, making it more important to correctly assess its legal status internally.
AI-generated code: where does the law stand?
As a rule, copyright law carries the high risk that the AI-generated output (e.g. a generated image) may not qualify for copyright protection. This is particularly relevant for cases with minimal human input, such as a one-line prompts.
This is because, according to the European Court of Justice (“CJEU”), copyright requires that a work “(…) constitutes an intellectual creation reflecting the freedom of choice and personality of its author” (see CJEU, Cofemel, C-683/17). This is not the case where the AI makes all or most of the defining creative decisions of a work.
The situation is rather nuanced in software development. While the use of GenAI may remain assistive to human work in different use cases (such as auto-suggestions), copyright eligibility remains a case-by-case assessment. Companies can only rely on arguments for protection if the human developer retains a dominant creative role. In such cases, developers typically work with substantial bodies of human-authored code[SIFH1] , which they amend, refine, or correct using GenAI tools. These human inputs can include generating short code snippets via prompts, identifying bugs, or optimising performance within existing codebases. Although binding case law at the EU or German level is still lacking on the eligibility to copyright protection of GenAI-generated code, companies could argue that:
- The original copyright protection of the code remains intact despite the use of GenAI, or
- Code produced with GenAI assistance, when primarily shaped by human input, such as an existing corpus of code and specific guidance, may meet the threshold of protectability under the standards of the European Court of Justice (see Cofemel, C-683/17).
Consequently, the company may assert that the human developer retained a decisive creative role: providing key inputs, contributing complex human-authored context, using GenAI merely to fill gaps rather than generate content from scratch, curating its suggestions, and integrating the results into a coherent whole. However, even in these assistive scenarios, determining where human creation ends and AI generation begins is complex and fact-specific.
However, where GenAI operates more autonomously and these conditions are not met, we end up with a lot of legal uncertainty around copyright protection.
AI disruption: how does agentic AI change this situation?
This situation is now being fundamentally challenged by the rise of agentic AI. Unlike purely assistive systems, agentic AI functions as an autonomous agent, capable of independently executing complex, multi-stage tasks with minimal human input. As a result, the human role shifts from direct creative engagement with the code, to simply defining high-level goals and framework conditions.
In specific use cases, this marks a transition from AI-assisted development to fully autonomous software generation. When agentic AI independently produces entire software modules - or even complete applications - based solely on general requirements and without meaningful human input, the level of human creativity needed for copyright protection is no longer met. There are only exceptional cases where you could still arguably assume a dominant human influence. For example, where the definition of the agentic AI's objectives includes extensive context (e.g. existing code) or detailed prompts, so that its use closely resembles that of conventional GenAI. However, this would in turn run counter to the very purpose of deploying an (autonomously acting) agentic AI, meaning these cases are likely to stay exactly that - exceptional.
According to the CJEU, a work must be an “intellectual creation reflecting the freedom of choice and personality of its author” to qualify for protection. Content created autonomously by agentic AI, lacking any human influence, may generally fall short of this threshold and enter the public domain.
For technology-focused companies, the use of agentic AI for code generation may raise critical concerns. Without transferable rights of use in the code, licensing conditions in the market could change drastically. Clients, in turn, might simply use an agentic AI to generate similar code themselves.
The widespread use of agentic AI in software development requires a strategic shift:
- If the generated software is intended to remain a core asset but is unlikely to qualify for copyright protection, other legal protections, such as confidentiality, become all the more critical. Alternatively, core IP must still be, at most, AI-assisted, but ultimately human-made.
- If the aim is to use agentic AI as a tool, rather than creating business-critical protected and licensable IP, then the value lies in the service layer. The focus needs to be on improving retention, scalability, and process efficiency. In these cases, the business model may shift from traditional licensing to offering proprietary know-how or superior agentic AI, tailored to specific use cases. The value proposition would lie in delivering results that exceed what clients could achieve on their own using agentic AI. This also raises questions of civil liability, technology regulation, and unfair competition law to the forefront of agentic AI’s actions.
What to consider next:
- Assess IP exposure from agentic AI outputs.
- Identify where agentic AI creates outputs with no or limited human input and analyse the legal and commercial impact if these are unprotected.
Navigating the AI Act: Classification challenges and regulatory complexity
In the context of agentic AI solutions, especially those encompassing autonomous software development, the classification and regulatory implications under the AI Act become increasingly complex. Initially, as with non-agentic GenAI use cases, companies face the core question of how individual AI models and systems should be classified under the AI Act. Specifically, a pivotal question is whether these models qualify as General Purpose AI (GPAI) models, as defined by the AI Act. Potential model modifications such as fine-tuning make this even more complicated. A critical consideration is whether such alterations at the model level result in the creation of new GPAI models, triggering additional regulatory requirements. Equally important is determining at what stage and how those models become formally integrated into a usable "system”.
These classification and integration questions can no longer be postponed, given that relevant GPAI provisions of the AI Act took effect in August 2025. In summer 2025, a substantial volume of official guidance documents were published, namely a GPAI Code of Practice, GPAI Guidelines, and template training data summaries. These documents offer practical guidance for companies navigating compliance with the AI Act. However, paradoxically, these guidelines sometimes introduce added layers of complexity.
For genuinely agentic systems, these challenges intensify. These systems often run not just through a single LLM, but through an ecosystem including multiple expert systems collaboratively tackling software development tasks. The interconnected nature of these AI ecosystems demands even more meticulous and complex regulatory analysis. Accurately establishing the roles, responsibilities, and classifications of each subsystem or model in this collaborative environment becomes both increasingly critical and notably more challenging under the AI Act.