Agents are useful mainly because they increase speed and scale, but their limitations make careful control necessary.
2
Curated documentation, smaller contexts, simple implementations, and multiple data representations can produce better agent results than adding more complexity.
3
You should not automate a job until you understand how to do it yourself, because evaluation depends on knowing what good work looks like.
Summary
Angus J. McLean argues that developers should place stronger boundaries around large language models instead of giving them more tools and more context. At Oliver, agents support advertising research, ideation, copywriting, content production, and performance work at high volume. McLean says the main value is speed, followed by scale, but models still learn inefficiently, fail to recognize new information, and do not continuously learn like people. Larger context windows help long-running tasks, yet they also introduce noise. His practical advice is to replace open internet access with curated documentation, reduce context until the task still works, build simple harnesses and memory systems, and test older models. He uses his own CV project as a warning: a complex application performed far worse than four letters, HTML. He also describes AI as translation between representations and recommends choosing structures such as Markdown, graphs, clusters, folders, or timelines according to the task.
McLean says Oliver uses agents primarily for speed and secondarily for scale. Creative teams use them to produce content quickly, especially when they need to iterate and test ideas. Strategy teams use them to expand research across audiences, territories, competitors, and trends. One example is localizing advertising research for cities such as New York or Miami. Oliver generates around 4,000 assets a day for more than 200 brands and puts substantial media spend behind many of those assets, which creates performance feedback from real campaigns. McLean also warns that higher volume can increase harm when a brand receives a poor response to its work.
McLean describes language models as closed boxes containing knowledge, with the ability to perform what he calls semantic math. He says they do not actually understand the data presented to them and do not continuously learn without forgetting as humans do. Models need massive datasets to reach conclusions that people can learn from a few examples. New trends are a practical failure point in advertising because a model may not recognize information that has only just appeared. McLean suggests that some recent progress comes from brute-force improvements to models and the compute used to train or run them, rather than from a complete change in their underlying nature.
McLean connects recent improvements in agentic systems to larger context windows. Long-running agents need room for goals, plans, action histories, tool outputs, and information they must retrieve during a task. Without enough context, a system can forget what it was doing in the middle of a complex workflow. However, context windows keep growing while the total amount of available knowledge also grows, so more capacity will not solve the problem by itself. McLean says the developer's challenge has shifted from getting enough context into the model to keeping irrelevant material out. He treats context as a soft constraint that can be shaped, rather than only as a limit to overcome.
McLean recommends withholding general internet access from a model when a task can be supported by high-quality documentation. In Oliver's advertising work, models are poor at spotting promotional material. When they research competitors, they may absorb the companies' own marketing instead of finding information from consumers. They are also susceptible to search-engine optimization. Giving the model a narrower tool changes what it can retrieve and how it uses that information. McLean contrasts this with older approaches from small-context systems, such as TF-IDF cluster labels and top-k retrieval. Modern context assembly is more dynamic, so he says the focus should be on excluding noise.
McLean argues that constraints can create better work because abundance can stop people from being scrappy. He asks how little of a model's context window is needed to complete a task, rather than assuming the whole window should be used. He compares this with early computing, when programmers had to achieve more with less memory and compute. He mentions Spacewar being built with only 4,000 words and developers of Crash Bandicoot making large improvements through careful use of PlayStation 2 memory. He suggests experimenting with older, smaller models or harnesses, building custom memory and compaction, and learning how preprocessing, archiving, file systems, and knowledge graphs affect control.
Simple implementations can outperform elaborate systems
McLean describes building a complex application to generate his CV, then discovering that the model produced a much better result when he simply asked it to use HTML. The four letters produced more than a 10x improvement, which he estimates may have been closer to 100x. He says models naturally produce verbose and complicated solutions, so giving them more power does not mean that developers should use all of it. A simple version that works gives a shorter feedback loop with reality. This is different from only shortening the feedback loop inside an agent workflow. The practical test is whether the implementation works in the real product or task.
AI systems translate between different representations
McLean presents AI as a translation process. He starts with the Transformer paper's English-to-French example, then extends the idea to text becoming images, images becoming audio, and audio becoming video. He also describes knowledge production as a form of summarization, where experience is compacted into a useful representation. Structured and unstructured forms can be converted in either direction, and a representation determines what a person can see or manipulate in the data. He relates this idea to model handoffs and MCP in long-running agent systems. The same content might be turned into a diagram, written explanation, or voice-over depending on the task.
Choose data structures according to how the agent will use them
McLean recommends using more than one representation structure when building agent systems. Markdown can provide a human-readable hierarchy and support authoring. Graphs can store relationships and references. Clustering can help organize large or unstructured bodies of text. Folders can make frequently retrieved material easy to find, while timelines may fit tasks that depend on sequence. The point is to decide what form makes the information useful for the operation being performed. McLean closes with a practical rule: do not automate a job unless you can do it yourself. He shows a social media intelligence report that clusters 50,000 tweets and organizes them into strategies, standout tweets, and other material for creative teams.
