Schaun Wheeler

I've never really enjoyed Wodehouse. I tried several of his books and thought they were all kinda meh. But I ate up everything from Deeping, Rinehart, Chambers, Morris, Oppenheim, etc. I don’t get Wodehouse’s appeal, given his contemporaries...which makes me feel I'm missing something important.

Question (honestly curious, not trying to be snarky): what do you find so perfectly executed about that story? I mean, it's delightful...but seems to be so in the same way as others of the same milieu, and with pacing a bit more stodgy than fits the character/setting.

An agent that can’t *choose* its next move isn’t an agent. It’s just a novel interface for the same information retrieval, content management, and marketing automation systems we’ve had for years.

Fully agentic systems need a hybrid architecture: a semantic–associative learner that builds and updates long-term user profiles, and a procedural actor that generates fluent, on-brand content (or retrieves it from inventory).

Such hacks don’t build conceptual understanding or learn from outcomes. If a user ignores a “20% off” push yesterday, an LLM can draft today’s new message about your return policy - but it won’t autonomously pick that message. No adaptation, no evolving preferences.

Recent LLM hacks like retrieval-augmented generation (external info stuffed into the prompt) or session-summary “memory” (re-feeding past interactions) preserve surface continuity, sometimes reduce hallucinations. But they aren’t true semantic/associative memory.

Real autonomy requires semantic + associative learning. An agent must consolidate experiences into transferable categories and tie those to outcomes. That’s how it forms opinions on which strategies to pursue or avoid over time.

Agency ≠ next-token prediction. A truly agentic system decides *when* and *how* to act without waiting for instructions. LLMs only predict the next token given a prompt. They don’t decide to prompt themselves.

Imagine someone who follows every cooking step flawlessly but has no sense of taste, no clue if others liked it, no idea how to improve. Without semantic understanding or feedback associations, true adaptation - and true agency - can’t happen.

Those are crucial to human cognition, but we also rely on:
3. Semantic memory = abstract concepts (knowing that “sustainability” is a thing).
4. Associative learning = linking concepts to outcomes (learning that stressing sustainability drives engagement).

LLMs excel at two kinds of “thinking”:
1. Procedural memory = automating skills (like writing a sentence or riding a bike).
2. Working memory = juggling info in the moment (like keeping a phone number in mind).

Procedural Memory Is Not All You Need: Bridging Cognitive Gaps in LLM-Based Agents Large Language Models (LLMs) represent a landmark achievement in Artificial Intelligence (AI), demonstrating unprecedented proficiency in procedural tasks such as text generation, code completion, and...

An LLM, by itself, cannot be truly agentic. Same for swarms, teams, workflows, or “multi-agent” systems. If the LLM drives everything, it’s not agentic. LLMs can be useful appendages, but not a sound foundation.

arxiv.org/abs/2505.03434

It’s hard to move beyond campaigns because they’re simple. They tame messy behavior into tidy segments. But simplicity for you isn’t value for users. An agentic mindset means letting agents manage orchestration’s complexity while analysis stays clear and human-scale.

Agentic systems separate orchestration from analysis. Orchestration is about maximizing who could benefit. Analysis is about retrospective learning—what worked, for whom, under which conditions. That separation expands impact without giving up interpretability.

That dual role makes campaign logic feel natural even when it’s arbitrary. Consider: “nudge users who haven’t engaged in 30 days.” Why 30? Why not 29, or 1? The threshold isn’t about user needs. It’s a simplification shaped by campaign design, not by actual behavior.

Campaigns usually serve two roles at once. First, orchestration: deciding which users get which messages under what conditions, breaking logistics into parts. Second, analysis: measuring outcomes tied to audience, timing, and content. Doing both jobs obscures insight.

A comparison of campaign-based engagement, which is easier to manage, with agent-based engagement, which is more effective because it decouples orchestration from analytics.

Technology is enabling, but also constraining. Choosing a tool means trading flexibility in one dimension for scale in another. Customer engagement is no different. Campaigns, in particular, aren’t neutral abstractions—they’re design choices with real consequences.

We don’t need to solve the philosophy of agency. What we need is a performance definition of acting agentically under complexity. Current benchmarks rarely test this, or when they do, they show LLMs fall short. That gap matters more than whether “agency” is solved.

Over minutes, mimicry looks convincing. But over hours, days, or weeks, acting agentically means deciding what to do next, why, and how to carry those lessons forward. For that, semantic–associative learning is required. Procedural memory alone isn’t enough.

When signals are delayed, goals conflict, or feedback is ambiguous, procedural mimicry fails. Without semantic memory to form abstractions and associative learning to link them to outcomes, systems can’t adapt with consistent success across shifting contexts.

In predictable, stable environments with clear feedback, even LLMs can appear agentic. With only procedural and working memory, they give the impression of knowing what they’re doing. But the appearance fades when environments become less structured.

In practice, the challenge is building systems whose behavior is hard to distinguish from beings who think for themselves. Acting agentically is a generalized Turing Test: not proving thought, but performing well enough that it looks like intention is present.

Because the concept of agency is so unsettled, I think it’s better to sidestep. A system doesn’t need to “have agency” in order to “act agentically.” That distinction matters more than trying to solve the philosophical problem of what agency really is.

We throw around terms like “agency” in AI, but the word itself lacks definition. That problem predates AI. Philosophers have debated it for centuries without consensus: Do instincts count? Are reasons different from causes? Is planning the same as action? Does coercion erase it?

95% failing sounds too positive to me. Traditional ML projects have long been pegged at 70–80% failure, after over a decade of playbooks, best practices, and data science professionalization. I’d have expected generative AI’s early-stage failure rate to be much higher than what that report shows.

If a system doesn’t do these six things, I have a hard time seeing how it could operate agentically. Large inventory, abstraction, broad rewards, individual learning, distributions, and priors aren’t really optional. They’re prerequisites for agentic behavior.

6. Can the system fill gaps in individual histories with inferences from others? Sparse data is unavoidable. Agents should be able to draw on similar users and impute where needed. Without this, they’ll fail early in a user’s journey, before there’s enough evidence.

5. Are you modeling performance as a distribution, not a point estimate? Confidence matters. One good outcome isn’t the same as ten. And outcomes should be weighted by probability, not just existence. Without this, I don’t see how agents can reason about risk.

4. Does the system track individual performance first? To be agentic, it has to learn how different message attributes work for each user, not just the average user. Aggregates are useful, but the agent’s view should be long, not wide. That feels essential to me.

3. Does the reward function let the agent use every behavioral signal? Not just clicks or conversions, but any sign the user is closer to a meaningful goal. If it can’t see that, the agent won’t have enough signal to calibrate its choices. That seems like a fatal gap to me.