Deterministic beneath Stochastic

If you’re building agentic systems, it’s tempting to use stochastic AI for all the steps in your process. For example among agent memory systems, letta works by trusting an LLM to summarize and organize facts, while systems like mem0 and graphiti use old fashioned techniques like NLP and graph databases.

But when the system doesn’t work the way you want, it’s much harder to reason about a system if it doesn’t have a bright line between the deterministic, explicit reasoning portion of the system, and the stochastic, inexplicable neural-net powered portion. Because when you start layering stochastic process derived data below your deterministic processes, you can’t track your error bars at a component level any more. When you have something implemented purely as an LLM prompt, the impact isn’t limited to your prose prototype. You have to treat the whole system as probabilistic.

Having non-deterministic layers makes engineering nearly impossible: you can no longer prove a fix truly removed a failure. It means that you can’t rely on data fidelity; large texts may get summarized totally differently between runs. It means the worst-cases and error/failure states may be way outside the caller’s expectations. It means even success may have completely different performance if you swap the LLMs powering a layer. It’s like a black box with problems you may never know about until you hit it in prod.

But it’s not just memory systems where I’m seeing folks leaning into determinism. There are software development lifecycle systems adopting tools like type systems, design-by-contract and property-based testing for deterministic validation. Production operations are building systems with DAG workflows, and giving broad read access to observability systems. And systems that need to provide access to existing data are leaning into semantic search for documents and SQL for structured data, even in places that used to use more custom query systems a few years ago.

A prose prototype may get you started fast, but switch to deterministic systems when the system needs audibility, traceable reasoning and predictable behavior.