Bounded Autonomy: Between Free Will and Determinism — Angus J. McLean, Oliver

Angus McLean, AI Director at Oliver — a generative AI-native advertising agency with 3,000 staff across 46 countries producing roughly 4,000 creative assets per day for over 200 brands — delivers a pragmatic talk challenging the default instinct toward complexity when building LLM-powered agents. The talk is framed around the concept of 'bounded autonomy': deliberately constraining what models can access and do, rather than maximising their degrees of freedom.

McLean opens by characterising LLMs as closed boxes capable of semantic computation rather than genuine understanding. He notes that models do not learn continuously, are susceptible to SEO-optimised content when given internet access, and struggle with recency — particularly a problem for Oliver since trend identification for advertising is core to their strategy work. He draws a parallel to historical computing constraints, arguing that tight resource limits have historically produced better engineering rather than worse.

A central claim is that most current agentic tooling constitutes 'band-aids' — temporary patches around fundamental model limitations rather than structural solutions. He singles out expanded context windows as the primary driver of recent agentic capability gains, enabling longer-running tasks by allowing agents to store history, tool outputs, goals, and plans in-context. However, he warns that the challenge has shifted from getting enough context in to keeping noise out: the question should not be 'how much context can I use?' but 'how little context do I need to complete this task?'

To illustrate the cost of over-engineering, McLean describes building a complex multi-step agent pipeline to generate his CV — only to discover that simply prompting the model to output HTML produced a result he estimates at 100x better. This anecdote anchors his principle of simplicity: models naturally gravitate toward verbose, complex outputs, and developers mirror that tendency. The corrective is to build the simplest version that works and shorten the feedback loop with reality before expanding scope.

On context management, McLean advocates replacing broad internet access with curated, high-quality documentation. In Oliver's workflows this matters because models are poor at filtering promotional content from organic consumer signals — a critical distinction when doing competitor research for advertising campaigns. He previously relied on TF-IDF and top-K methods for cluster labelling when context windows were small (e.g., 512 tokens); today context assembly is more dynamic, but that abundance introduces its own noise problem.

McLean frames AI broadly as a translation process — citing the original 'Attention Is All You Need' English-to-French formulation and extending it to cross-modal translation (text to image, image to audio, audio to video). He argues that knowledge production itself is a form of summarisation and that if an object can be manipulated through a representation space, its structure is a property of the observer's chosen representation rather than the object itself. He recommends using multiple representation structures in agent memory: markdown for human-readable hierarchy, graph relationships for references, clustering for large unstructured corpora, folder hierarchies for fast retrieval, and timelines where temporal ordering matters.

Practical advice for developers includes: building your own harness and memory compaction rather than relying entirely on framework abstractions; experimenting with older or smaller model versions to build intuition; and pre-processing and archiving data carefully. A closing rule — 'don't automate a job you can't do yourself' — is illustrated with a social media intelligence report he built over 50,000 clustered tweets, producing segmented strategic insight for creative teams. He contends that workflows are consistently more effective than unconstrained agents, drawing an analogy to Adam Smith's pin factory and task decomposition.