AI Engineer Europe 2026
Bounded Autonomy: Between Free Will and Determinism
Loading slides…
Read the talk
Bounded Autonomy: Between Free Will and Determinism
Advertising agents need speed, but useful autonomy depends on selected context, simple workflows and representations that keep the work understandable.
From a talk by Angus J. McLean
Advertising that has to work in the wild
How much freedom should an agent have when its output reaches real customers? That question sits between automation and customization, oversight and agency, possibilities and constraints. Angus J. McLean approaches it through experience designing agents in advertising: practical guidance for beginners overwhelmed by change and experienced builders looking for another perspective, rather than a definitive prescription.
McLean introduces himself as AI director at OLIVER, an advertising business that has shifted heavily toward generative AI. He reports 3,000 staff across 46 countries. The Johnnie Walker work provides the concrete starting point: an audience can encounter an AI-generated advertisement without recognizing how it was made. The business problem is therefore larger than generating an impressive image. The image has to function as advertising.
McLean reports around 4,000 assets a day for more than 200 brands, with media spend ranging from twenty thousand to a few million; he does not specify the currency. Paid distribution makes the assets observable in the market. Performance data comes back from real exposure, allowing teams to iterate and develop a better understanding of what works. Generation becomes useful through its feedback loop.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Speed first, scale second
An agency is easier to understand through its responsibilities than through the familiar Mad Men image. Accounts manages client relationships and keeps projects on track. Creative turns ideas into ads and content. Strategy establishes the insight, audience and direction behind that work. McLean describes accounts as historically about 50% of an agency, alongside 20% creative and 20% strategy; these are a rough sketch, with mixed time references and no allocation for the remainder.
Creative and strategy are increasingly places where agents do knowledge work:
- Creative: ideation, copywriting and content production.
- Strategy: audience insight, trend analysis, competitor analysis and performance optimization.
McLean works on the strategy side. These applications face customers, so their speed carries risk: scaling content can amplify a poor reception just as readily as a successful campaign.
The primary reason to use agents is speed; scale comes second. Creative teams need to generate, test and revise content quickly. Strategy teams need to expand research so they can understand consumers more closely and support eventual conversion. Campaign personalization connects these tasks: define the relevant personas or territories, research each one, and use that research to shape the work. Localizing advertising for New York and Miami, for example, requires more than producing extra copies of the same asset. It requires understanding each location. The goal is more effective advertising with fewer resources—which leads to McLean's counterintuitive first recommendation: slow down.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Slow down enough to see what persists
Tool churn creates pressure to keep up with everything. McLean asks, “If you did blink and you did miss it, was it really that important?” His underlying argument is that durable limitations deserve more attention than short-lived tools. He places the foundations of LLMs at least as far back as the 1990s, invoking Emma Strubell as someone who would reach further back, and argues that apparent sophistication should not be mistaken for understanding. These are his conceptual premises for designing agents.
Two limitations matter especially to that design. Data efficiency concerns how much experience learning requires: humans can learn from few examples, while models depend on large datasets. Continual learning concerns whether a system can keep learning without forgetting. McLean contrasts human learning with a model whose trained knowledge remains closed during use. He attributes some recent gains to brute-force compute, invokes The Bitter Lesson, and recalls a previous presentation's image-generation analogy of 400 marathons. That analogy has no defined workload or conversion here, so it cannot establish a compute cost.
Against those persistent constraints, a tool can become a temporary patch. McLean's test is whether it is short-lived, superficial, masks symptoms or only partly resolves the problem. The useful engineering question is what a surrounding system actually fixes—and what limitation it merely makes less visible.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A closed model needs working context
McLean's simplifying model is a closed box of knowledge: a flexible database capable of “semantic math.” He does not design around an expectation that it will spontaneously learn or develop new capabilities. Advertising makes the limitation concrete. A genuinely new trend may be absent from the model's knowledge, precisely when a strategist most needs to understand it.
Larger context windows let an agent work with more information during a task. McLean credits them with much of the progress in long-running agents because they can hold the action history, tool outputs, task organization, goals and plans together. That context supplies short-term storage and retrieval: the agent can inspect what it has already done and decide what comes next. When relevant information falls out, a multistep workflow can lose continuity. McLean invokes an unspecified email-deletion incident as a warning about context exhaustion, without enough incident detail to establish its cause.
The recording compares GPT-2 with a model named aloud as Gemini 3.5 Pro, then recalls 512-sized windows. Those details do not establish a reliable capacity comparison: GPT-2's implementation defaults to n_ctx=1024, and the spoken Gemini identifier remains uncertain. The organizer's schedule places the talk before Google's announcement, which still described 3.5 Pro as forthcoming. The broader design point is finite working space. McLean argues that demand will always outgrow it, illustrating that argument with an unsupported claim that world knowledge doubles every 12 hours. The practical problem does not require that growth rate: a larger window still requires decisions about what belongs inside it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Constrain the evidence, then exclude the noise
Context is a soft constraint. What an agent receives, and how that information is arranged, steers its behavior without enforcing a hard limit. McLean's everyday example is to replace broad Internet access with high-quality documentation. The distinction matters: restricting access is a capability boundary; choosing the documentation shapes the evidence available inside that boundary.
Competitor research exposes the failure mode. An agent looking for consumer insight can absorb the competitor's own promotional language instead. Search visibility does not make a source representative of consumers, and McLean finds models susceptible to SEO and poor at recognizing promotional material. Once knowledge is retrieved through a tool, the result depends on how the agent uses that tool as well as on what the model knows. Selecting sources is therefore part of controlling the task.
Earlier, with small windows, McLean used TF-IDF for cluster labeling. The procedure was to cluster a text corpus, extract characteristic words from each cluster, and use those words to produce a label. TF-IDF weights word frequency by how distinctive a term is across documents, rather than simply counting the most common words. Top-K selection similarly limits how much material enters the context. With larger windows, assembly can be more dynamic, but the problem changes: getting information in becomes easier, while keeping irrelevant information out becomes more important. The target is useful context, not a full context window.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Practice with less
If model progress stopped tomorrow, what could you still build? That question turns scarcity into a design exercise. Instead of asking how much context a model supports, ask how little context completes the task. McLean jokes that he is not yet a “token billionaire,” then draws a parallel with early computing and the Model Railroad Club. He invokes Spacewar! as something built with only 4,000 words, though the talk does not establish whether that figure describes program size or available memory.
The next analogy is Crash Bandicoot's memory optimization. McLean says PS2, but the corroborated development account from Andy Gavin concerns the original PlayStation. Its relevance is the engineering response to limited memory: decide what must be resident and what can be brought in when needed. McLean carries that spirit into suggested experiments, explicitly separating them from production recommendations.
The exercises move progressively closer to the system's mechanics:
- Try an older or smaller model, or a simpler harness, to make its limitations easier to observe.
- Build a harness and implement memory and compaction, so that decisions about retaining and discarding information become explicit.
- Work on preprocessing and archiving; study how file systems and knowledge graphs organize information for later use.
The intended benefit is better prompting and control, backed by a closer understanding of the data. McLean ends this sequence with an image he describes as Rosenblatt pulling wires from a perceptron and connects it to dropout. It is a suggestive analogy to removing connections, rather than an established historical lineage.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Shorten the feedback loop with reality
Sometimes the most useful constraint is to remove the application you thought you needed. McLean describes building a complex CV application that worked, but worked less well than asking for the result in HTML. McLean estimates that the HTML approach improved his CV task by more than 10×, perhaps 100×, without defining a measured metric. The useful comparison is the implementation choice: a direct document format solved his problem better than the surrounding machinery.
Models can make unnecessary complexity easy to produce. McLean characterizes them as verbose and inclined to suggest complicated solutions, leaving the builder with more tokens spent and more work to maintain. Build a simple version that works. The same short feedback loop that makes agents attractive should apply to product development: get the idea into contact with reality quickly enough to discover whether the extra machinery helps.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Choose the representation that serves the task
McLean next treats AI as translation. Attention Is All You Need supplies the starting example: machine translation, including both English-to-German and English-to-French experiments. The broader analogy extends beyond language pairs to text becoming images, images becoming audio, and audio becoming video. Knowledge production, in his framing, is also summarization: this talk compresses years of experience into a smaller body of usable knowledge.
The practical mechanism is transformation between representations. Unstructured input can become structured output, and structured input can become prose or another less structured form. McLean connects this idea to MCP and handoffs in long-running agents. The analogy should remain distinct from the protocol's scope: the later Model Context Protocol documentation defines connections between AI applications and external data, tools and workflows, rather than a general model-to-model handoff mechanism.
If information can move between forms, its useful structure depends partly on what the observer needs to do with it. Slide design makes that tangible. Some material belongs in a diagram; some needs written explanation; some is better voiced over. A single preferred format can obscure these differences. Choosing a representation is part of solving the task.
That leads to several complementary structures:
| Representation | Useful for |
|---|---|
| Markdown or outlines | Readable hierarchy and authoring |
| Graphs | Relationships and references |
| Clusters | Large or unstructured text collections |
| Folders | Fast retrieval through known locations |
| Timelines | Questions where sequence matters |
For example, a Markdown research note can make the distinction between a finding and its evidence explicit:
markdown
# Audience research
## Finding
State the consumer insight.
## Supporting evidence
Record the source and the relevant observation.
## Implication
Describe what this changes in the creative brief.
This is an authoring structure, not a completed analysis. The same underlying material can also participate in a graph of relationships, a cluster of similar observations or a timeline. Multiple representations let different tasks use the same information without forcing every task through one format.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From experimentation to a usable intelligence report
Some of this understanding comes from thoughtful play. McLean reports attending more hackathons than before because they permit experiments he cannot try within his job. Work imposes necessary boundaries; experimentation outside those boundaries creates room to learn techniques that can later improve professional practice.
The closing question is how much structure agents need in the workplace. McLean reaches back to Adam Smith's pin factory: work can be decomposed into small, repeatable tasks. In his experience, workflows have often proved more effective. That makes domain competence a prerequisite for useful automation. His rule is direct: “don't automate a job unless you can do it yourself.” Knowing the work gives the builder a basis for deciding how to divide it and whether its output is useful.
The final demonstration returns to McLean's former job in social media intelligence. He shows an LLM-produced report about OpenClaw, using roughly the preceding year's data. The model has clustered the material and assembled the kind of report an advertising agency would use.
McLean reports that the OpenClaw analysis clusters 50,000 tweets and organizes them into strategies, describing report creation as almost instantaneous without giving a measured runtime. The report also includes standout tweets and a breakdown of the insights. Its destination is a creative or strategy team: a large body of social conversation has been organized into something people can use to decide what to make next. That is where the closing demonstration locates the value of the agent—in the passage from raw material to useful work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The Transformer paper, reporting machine translation experiments in English-to-German and English-to-French.
GPT-2's implementation, including its default 1,024-token context setting.
The personal AI assistant project discussed as the subject of the closing social-intelligence report.
Further reading
A developer's account of fitting PlayStation game levels into limited memory using streaming, compression and custom tools.
The 2014 paper explaining dropout as random removal of units during training to reduce overfitting.
Updates since the talk
Later documentation explaining how MCP connects AI applications to external data sources, tools and workflows.
Full transcript
Loading transcript…