Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Bounded Autonomy: Between Free Will and Determinism — Angus J. McLean, Oliver
TL;DR
Angus McLean’s core advice is to slow down and design around limits — despite the hype cycle, he argues the “core” of LLM behavior hasn’t fundamentally changed, and most agent tooling is still a temporary band-aid around model constraints like context and retrieval.
At Oliver, agents are already doing real commercial work at scale — the company says it generates roughly 4,000 assets a day for 200+ brands across 46 countries, often with $20,000 to multi-million-dollar media spend behind them, which creates a rare performance feedback loop in the wild.
Context is doing more than guardrails ever will — McLean frames context as a soft constraint that shapes model behavior, saying a model with curated documentation and no internet often outperforms one let loose on the web, especially in ad research where SEO and self-promotional content distort results.
The bottleneck has flipped from ‘not enough context’ to ‘too much noise’ — earlier workflows relied on tricks like TF-IDF and top-K labeling because windows were tiny, while today the harder problem is deciding what to exclude so agents stay focused and don’t drown in irrelevant tokens.
His strongest product-building lesson is brutally simple: keep it simple — he recounts building an overly elaborate CV app only to get a 10x to 100x better result from “four simple letters,” HTML, as a reminder that models naturally drift toward needless complexity.
He sees AI fundamentally as translation across representations — from the original English-to-French framing of Attention Is All You Need to text-image-audio-video workflows, he argues builders should think in multiple structures at once: markdown, graphs, clusters, folders, and timelines.
The Breakdown
Why this talk exists: conventional wisdom for agent builders
Angus McLean opens by positioning the talk as practical advice for people actually building agents, not a technical manifesto. He’s AI Director at Oliver, a company that shifted from advertising into “almost fully gen AI,” and he frames the whole session as a perspective reset for beginners overwhelmed by change and experts who feel stuck.
What agents look like inside a modern ad agency
He gives a quick anatomy of agency work: accounts, creative, and strategy — with creative and strategy now becoming increasingly agentic. At Oliver, that means agents handling ideation, copywriting, content production, audience insight, competitor analysis, and performance optimization, all in a fast-moving, high-risk environment where bad outputs can damage a brand just as easily as good ones can help.
The real-world scale: 4,000 assets a day and measurable feedback
This isn’t hypothetical experimentation. Oliver generates around 4,000 assets per day for more than 200 brands, and unlike many AI content shops, they actually put serious paid media behind the work — from about $20,000 to several million dollars — so they can measure what performs in the wild. That gives them the thing most AI demos don’t have: a live feedback loop tied to real consumer response.
Slow down: most of the industry is still wrapping band-aids around model limits
McLean’s first big recommendation is almost anti-hype: slow down. He argues current LLMs still don’t “understand” data in a human sense, and their weaknesses — poor data efficiency, brittle learning, closed-box behavior — mean a lot of the ecosystem is just patching around those constraints rather than solving them. His image for the whole stack is memorable: the tools are band-aids, temporary and superficial, often masking symptoms more than fixing causes.
Context windows made agents possible, but they’ll never be enough
He describes LLMs as a “closed box with knowledge inside,” more like a flexible database doing semantic math than something genuinely emergent. Bigger context windows have powered long-running agents because they let models retain plans, tool outputs, and action history, but he says they’re still fundamentally insufficient — especially in advertising, where fresh trends often appear before the model can recognize them. The danger is obvious: when context runs out, the system forgets mid-task, and weird failures follow.
Context as a design tool, not just a capacity limit
One of his sharper points is that context constrains a model almost as much as formal guardrails do. In practice, he says you often get better results by removing internet access and feeding in high-quality documentation, because models doing competitor research tend to absorb SEO sludge and brand self-promotion instead of real consumer sentiment. The builder’s challenge has changed from squeezing enough context in to aggressively keeping noise out.
Constraints create creativity, and simplicity beats god-mode engineering
He makes a broader engineering argument: self-imposed constraints produce better systems. He points to old computing culture, Spacewar, and even Crash Bandicoot memory hacks as proof that great work comes from limitations, then suggests experimenting with smaller models, custom memory, compaction, preprocessing, and better file or knowledge structures. The emotional high point is his own humbling story: he built a complicated CV app, then got a dramatically better result — “probably 100x” — from plain HTML.
AI as translation, plus one last warning about automation
Near the end, he zooms out and says AI is fundamentally translation: English to French, text to image, image to audio, structured to unstructured, and back again. That leads to a practical design principle: use the representation that fits the task — markdown, graphs, clusters, folders, timelines — instead of forcing everything into one format. He closes with a grounded workplace note: workflows often outperform looser agents, and “don’t automate a job unless you can do it yourself,” illustrated with an example report built from 50,000 clustered tweets turned into near-instant strategic insight for creative teams.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
