Building SDKs in the Agentic Era — Mark McDonald at AI Engineer Melbourne 2026

The rapid evolution of software development has reached a critical juncture, particularly with the advent of frontier AI models and coding agents. Mark McDonald, a prominent figure in the AI and developer experience community, is set to delve into this shifting landscape at AI Engineer Melbourne 2026. He will explore the challenges faced by developers and SDK (Software Development Kit) maintainers in an era where traditional stability in SDKs is increasingly elusive.

Historically, SDKs provided a stable interface that developers could rely on for extended periods. Documentation was created, developers learned to navigate the SDK, and it remained relatively unchanged for years. However, this paradigm has shifted dramatically. In today’s environment, an SDK’s documentation may become outdated almost as soon as it is published. As McDonald notes, a large language model (LLM) trained on current SDK documentation may suggest coding patterns that were relevant in 2024 but are now considered outdated or inefficient by 2026. This discrepancy arises because the training of these models occurs on a timeline that spans months, while the evolution of APIs can happen in just weeks.

This rapid pace of change introduces a unique form of technical debt—not within the code itself, but in the disconnect between what AI systems are trained to understand and the current state of the software landscape. For teams that are tasked with delivering both innovative AI models and user-friendly SDKs, this gap is not merely a theoretical concern; it manifests as a daily operational challenge.

The Developer Experience Loop

To illustrate this issue, consider the developer experience loop. A data scientist may seek to integrate a new AI model into their application and turn to their preferred LLM for example code. The LLM, drawing from its training, may confidently suggest an API pattern that is based on outdated information. If the SDK has undergone changes in the preceding months—such as the removal of deprecated parameters—the suggested code may not only be inefficient but could potentially fail altogether. This situation breeds frustration among developers, who may feel let down by the SDK team, leading to a cycle of dissatisfaction.

The crux of the problem lies in the timing of LLM training versus the pace of API changes. While LLMs are powerful tools, they are limited by their training data, which becomes static at the point of training. In contrast, the open-source ecosystem is dynamic and responsive, which is beneficial until an AI agent attempts to navigate it with outdated knowledge.

In response to these challenges, the team at Google DeepMind has been conducting empirical experiments to quantify the knowledge gap between LLMs and current best practices. They are investigating the types of changes that are most likely to confuse AI agents, how often LLM suggestions align with up-to-date practices, and what SDK maintainers can do to enhance the legibility of their libraries for future AI systems.

The findings from these experiments have revealed some unexpected insights, as well as confirmations of existing beliefs. Breaking changes in APIs are indeed significant, but so too is the quality of documentation. A well-structured changelog that not only details what has changed but also explains the rationale behind those changes can aid models in understanding transitions. Providing examples that illustrate before-and-after scenarios gives AI agents a reference point, while consistency in naming conventions and structural organization allows for easier generalization by models.

Redefining SDK Design

This understanding necessitates a rethinking of how SDK design is approached. Developers must recognize that they are not only designing for human users but also for the LLMs that will generate code based on their libraries—both existing models and those yet to be developed. This does not imply sacrificing clarity for human users in favor of machine readability; instead, it calls for a more deliberate approach to structuring documentation, being explicit about versioning, and thoughtfully managing deprecation paths.

SDK maintainers should treat changelogs as integral components of API documentation. Consistency across similar functions is essential; if one function adheres to a pattern, both human and AI users will expect others to follow suit. Recognizing that an SDK serves not only as an interface for services but also as training data for AI systems is crucial.

The implications of these findings extend throughout the software development ecosystem. As more developers turn to AI agents for code generation, the SDKs that will thrive are those designed with AI compatibility in mind. This evolution is not about simplifying APIs or reducing functionality; rather, it is about intentionally exposing and documenting that functionality in a way that is accessible to both humans and AI.

Mark McDonald, who has co-authored the Guinness World Record-holding Kaggle Generative AI course and collaborates with DeepMind research teams, is at the forefront of this discourse. His experience with various APIs, including Gemini, PaLM, TensorFlow, and Google Maps, positions him uniquely to address the complexities of developing SDKs that align with the demands of this new era. McDonald will share insights and experimental results regarding the training cutoff knowledge gap at AI Engineer Melbourne 2026, scheduled for June 3-4, further illuminating these pressing issues in the field of AI and software development.