Descriptions:
Angus McLean, AI Director at Oliver—a generative AI advertising agency producing roughly 4,000 creative assets per day for over 200 brands including Johnnie Walker—presents a practitioner’s framework for thinking about LLM agent design in high-volume, high-stakes production environments. His central concept, “bounded autonomy,” is a mental model for calibrating how much freedom to give an agent versus how much structure to impose, drawing on hard-won lessons from running agents at advertising scale across 46 countries and 3,000 staff.
The most actionable section covers context management. McLean argues that context acts as a soft constraint on model behavior—not just a resource to fill—and that excluding noise is often more powerful than adding information. He shares a concrete lesson about keeping agents away from live internet access in favor of curated, high-quality documentation, noting that models are poor at filtering promotional content and highly susceptible to SEO-optimized pages when doing competitive research. He also traces the shift from context scarcity (TF-IDF cluster labeling for small windows) to context abundance, where the challenge is now quality filtering rather than information retrieval.
McLean situates these technical observations inside a broader organizational transformation: Oliver has restructured from the traditional accounts/creative/strategy agency model to one where creative and strategy functions are increasingly agentic. He explains why agents are deployed primarily for speed and secondarily for scale, and makes the counterintuitive case that slowing down to understand fundamental LLM behavior—which he argues has been stable since the 1990s—is itself a competitive advantage in an industry prone to chasing every new tool release.
📺 Source: AI Engineer · Published May 25, 2026
🏷️ Format: Deep Dive
