Notes from Vacation: Logical Positivism, LLMs, and Agents

Flying back from a vacation in Switzerland and Austria. I expected natural beauty and cities with grand palaces and churches – and on both counts the trip was spectacular. What I didn’t expect to come home with was a sense of Vienna’s extraordinary intellectual heritage. From the late 19th century until the mid-20th century, a remarkable community of philosophers, mathematicians and scientists gathered there and, together, produced a rich body of knowledge across disciplines. Reading more about this fascinating phase in this book by Richard Cockett.

Logical Positivism

One such idea, from the 1920s and 30s, was logical positivism. It was built on the verification principle: the claim that a statement is meaningful only if it is either true by definition (4 + 8 = 12) or empirically verifiable (testable by observation). Everything else — metaphysics, theology, much of ethics — was, on this view, not false but literally meaningless. This was the age of rational, scientific thought, and for a while it was an exhilarating idea.

It was also short-lived, and was challenged most notably by the Vienna-born philosopher Karl Popper. He pointed out that scientific laws like “all electrons have the same charge” can never be conclusively verified — no finite set of observations is ever enough. He offered falsifiability instead: what makes a claim scientific is that some observation could prove it wrong. But that too struggled as a theory of meaning. And there was a deeper, fatal problem: the verification principle could not pass its own test. It was neither true by definition nor verifiable by observation. By its own rule, it was meaningless.

Initially I thought this was an abstract debate for philosophers. But the more I dug, the more it struck me how practical and current it actually is — which is why I’m writing this post. (Sidebar: many of the questions philosophy wrestles with turn out to have very real, everyday counterparts. More on that another time.)

Why does this matter

The positivists wanted individual sentences to wear their empirical meaning on their sleeve. But as we know well in real life, statements rarely meet the world one at a time. Empirical meaning is not an isolated property of a sentence; it exists within a larger network of facts, events and prior assumptions.

This is exactly the right way to think about LLMs. We know by now that LLMs are superb at producing sentences that appear empirically meaningful — and they do it through statistical correlation, nothing more. That, in fact, is what “hallucinations” really are. Not some kind of bug in the model, but the natural behaviour of a system that operates in isolation, without the surrounding network that would confer meaning.

If meaning is conferred by a framework, then the framework has to be supplied from outside the model. This is the whole game in Enterprise AI. The raw model is signal; what turns its outputs into something evidence-like is the scaffolding we build around it: retrieval over verified sources, tools that actually query the world, citations a human can check, evaluation that closes the loop.

And it changes how we should use these systems. Fluency and confidence, without grounding, mean nothing. The Vienna Circle failed to find a rule that could certify meaning from the surface of a sentence; a century later, we are relearning the same lesson with our most powerful machines. The question worth sitting with is not “which model is smartest?” but “what is the framework that lets a model be useful in the world?” We know by now that it is not individual data points but the ‘context network’ which is a combination of assumptions (‘The markets a company will operate in’), baseline facts (‘network of manufacturing centers supplying a geography’) and temporally sensitive facts (‘latest inventory at each point in a supply network’). Each of these by themselves is of limited value, but when put together, can provide the right context for the models to be useful.

A layer above the model

On the last day of the trip I turned my laptop on and saw that Databricks had just open-sourced Omnigent. Enterprise agentic systems are already being built in layers — reasoning (model intelligence), tooling (governed connectors and actions), orchestration and durability (retries, approvals, systems of record), evaluation (the quality of an agent’s actions), and governance (identity, audit, policy enforcement).

Omnigent adds a layer above all of these: a meta-harness that sits on top of individual agent harnesses like Claude Code, Codex or your own custom agents. It is where composition, control and collaboration live — combining models and sub-agents, and governing them with stateful, contextual policies (cost budgets, permissions, guardrails) enforced at that layer rather than buried inside prompts.

The details differ from my retrieval-and-verification example, but the shape is the same. The intelligence sits in the model; the meaning, the control and the usefulness sit in the layer we wrap around it.

And so the philosophical questions remain open, even as the frameworks keep evolving to make agents meaningful and useful in the real world. The positivists looked for meaning inside the sentence and came up empty; we have quietly stopped looking there, and started building it into the layer around the model instead. That a failed idea from 1930s Vienna and an open-source release from 2026 San Francisco should rhyme, in the same week, is exactly the kind of serendipity that this blog aims to celebrate!