Angus McLean on Bounded Autonomy in AI

Angus J. McLean, an AI Director at Oliver, recently delivered a talk titled "Bounded Autonomy: Between Free Will and Determinism." The presentation, held at an AI Engineer Europe event, explored the complexities of AI development and its application, particularly within the advertising industry. McLean, whose company Oliver is a significant player in the generative AI production space with a global team of 3,000 across 46 countries, shared insights into how AI agents are transforming creative and strategic processes.

The Agentic Shift in Advertising

McLean highlighted that while many may not realize it, AI is already deeply embedded in the content they consume daily. Oliver, for instance, generates over 4,000 assets daily for more than 200 brands, putting millions of pounds into media campaigns each month. This scale of operation, he explained, allows for tighter feedback loops, faster iteration, and a deeper understanding of what truly resonates with audiences. He noted the shift from traditional advertising agency structures, which historically focused on accounts, creative, and strategy, to a more agentic approach where AI plays a central role in these functions.

Understanding and Limitations of AI Agents

The talk delved into the current state of AI, particularly large language models (LLMs). McLean posited that despite their advanced capabilities, LLMs still have significant limitations. He referenced the concept of "bounded autonomy," suggesting that while AI can perform complex tasks, their understanding is not equivalent to human cognition. Key limitations identified include a tendency towards verbosity, overestimation of their own abilities, and a susceptibility to issues like forgetting or being influenced by promotional context and SEO. He illustrated this with the example of how providing high-quality documentation, rather than simply web search access, yields better results, as models are not adept at discerning genuine promotional content.

The Role of Context and Constraints

A significant portion of McLean's presentation focused on the importance of context and constraints in managing AI agents. He differentiated between "soft constraints" (guidance derived from prompts, conversation history, and instructions) and "hard constraints" (rules, policies, and system-level restrictions that define what the model must not do). McLean emphasized that context engineering is crucial, especially with the increasing complexity of AI models and the need to manage the vast amount of data they process. He cited the evolution of LLM context windows, from 1,024 tokens in 2020 to 1 million tokens in models like Gemini 3.5 Pro, as a key driver of recent advancements in agentic capabilities.

Practical Advice for Building AI Agents

McLean offered several practical pieces of advice for those looking to build or work with AI agents:

• Slow Down: Given the rapid pace of AI development, it's important to approach new tools and techniques deliberately.
• Constrain Context: Effectively managing the information provided to AI models is crucial for achieving desired outcomes.
• Keep It Simple: Start with the simplest version of a solution that works and iterate from there.
• Meet The Model In The Middle: Understand the model's capabilities and limitations to find the most effective ways to collaborate.
• Have Fun & Experiment: Encourage exploration and learning, as many breakthroughs come from unexpected discoveries.

He also touched upon the idea that progress in AI is often driven by constraints, citing historical examples like the development of Spacewar! and Crash Bandicoot, which were limited by the computing power of their respective eras. This suggests that limitations can foster creativity and efficiency.

The Future of AI and Knowledge Production

McLean concluded by emphasizing that the core function of many AI models, at their heart, is translation, converting one form of data into another, whether text to image, text to audio, or text to video. He argued that knowledge production itself can be viewed as a form of summarization, where complex information is distilled into more digestible forms. The key, he suggested, is to understand that structure is not an inherent property of data but a property of its representation and the observer, reinforcing the importance of thoughtful context management and experimental approaches in harnessing the power of AI.