--- title: "Emerging Field of Epistemic Engineering or Stochastic Management" description: "Practices around context, uncertainty, and evals are hardening into a discipline, with belief states at its centre: probabilistic, grounded views of a work process." type: note pubDate: 2026-09-25T00:00:00.000Z topics: ["agents", "ai", "context", "uncertainty", "evals", "knowledge"] canonicalUrl: "https://actuallymaybe.com/blog/emerging-field-epistemic-engineering/" --- > Practices around context, uncertainty, and evals are hardening into a discipline, with belief states at its centre: probabilistic, grounded views of a work process. One thing I've been thinking about is the evolving nature of how we deploy agentic tools. Certain practices are beginning to develop around managing context and managing uncertainty. There was the viral launch of Jev, a decision model (not sure if that is the correct term).[^1] Don't you see that different components are being carved out into discrete little domains? My intuition tells me it's becoming a discrete engineering practice, something like epistemic engineering or stochastic engineering, with different areas of concern: quality control, evals, simulation of situations, and context management. And there's a new space, which I think should be characterised as belief states or belief networks. You're plugging it into a work process, and rather than canonically having an exact idea of what's happening, you always have a probabilistic understanding of the situation. For example, you could have a system that's linked into Slack, invoicing, and maybe Google Drive. It sees Slack messages from one person to another saying, "Oh, X has paid," and then there's a response that says, "You forgot this other piece." That's the entire conversation, and the two humans in the loop know the rough context of what they're talking about. But you want to have this idea, this concept, of belief states based on a certain grounding in reality. The building of those systems encapsulates what I think is epistemic engineering. How do you evaluate what's working and what's not working, the trade-offs and costs, plugging in new models, trying different techniques? There's a lot here. ## Related Thoughts - [The Different Problems of Searching Global and Local Context](/blog/global-vs-local-context-search/): the Slack example is a local-context problem. The two people share background that a system has to infer. - [The AI Tool That I Wish Existed](/blog/ai-tool-wish-existed/): an earlier sketch of a tool that sits across existing apps and quietly links everything together. - [Why Can't I Talk to My Tools, and Why Don't They Know What to Do?](/blog/quick-thought/): tools that understand the situation you're in, rather than waiting to be told. - [Exploring the Primitives of an Agentic Coding Setup](/blog/exploring-primitives-agentic-coding-setup/): the same instinct to carve agentic work into discrete, foundational pieces. - [Cheaper Code Should Mean More Experiments](/blog/cheaper-code-more-experiments/): cheap, throwaway experiments are how you find out which models and techniques actually work. [^1]: Jev is TypeSafe's "System One" model, launched in September 2026. It returns typed decisions with calibrated probabilities rather than free text. See [Introducing System One Models & Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev). ## Further reading - [Notes on Visual Interfaces, AI Tooling, Fast LLMs and Decision Models](https://actuallymaybe.com/blog/notes-visual-interfaces-ai-tooling-fast-llms-decision-models/) — Why agentic work is stuck in text, and what canvases, fast models, branching version graphs, and decision models might add. - [Building Explainers Random Thoughts - Drone 101](https://actuallymaybe.com/blog/building-explainers-random-thoughts-drone-101/) — An experiment in interactive explainers: drone control, Fourier, Bret Victor, philosophical prompts, accessibility, and learning with LLMs. - [Cheaper Code Should Mean More Experiments](https://actuallymaybe.com/blog/cheaper-code-more-experiments/) — When working with agentic tools, the instinct to solve hard problems inside your main project is usually wrong. Isolate the core problem first, experiment freely, then bring what works back. - [Software Architecture That Agents Actually Like](https://actuallymaybe.com/blog/agent-friendly-architecture/) — Structural patterns that make codebases easier for AI agents to navigate: composition over monoliths, stateless mini-CLIs, typed languages, and the editor navigation test. - [Using Agentic AI Tools for Anki Flashcard Creation](https://actuallymaybe.com/blog/agentic-ai-anki-flashcard-creation/) — How Claude Code and similar tools can handle the tedious work of expanding concepts into multiple flashcards, and why Anki Connect is the key to making it work.