Nondeterministic agents are the trend of 2026, which I am diving into up to my ears.

Nondeterministic agents are the trend of 2026, which I am diving into up to my ears.

What is a deterministic agent? It is a rigid system with a clear flow: like in n8n, Make, and any other systems where a clear action algorithm is predefined.

Advantages of such systems — they are easy to debug, put on rails, and relatively simple to assemble. Essentially, now anyone with a good head on their shoulders can go into Cursor / Claude Code / n8n and assemble their content generation service from a pipeline like: analysis → research → RAG → generation → editing.

But such systems also have disadvantages: little flexibility, many unnecessary steps, the user cannot properly add their own work scenario. Flexibility usually appears only because you create a bunch of pipelines for all possible life cases.

I am often asked (without bullshit): why do I give away my content generation scheme in n8n with all the prompts if it can easily be put into Cursor / Claude Code and a service can be made based on it? And why don't I make such a service myself?

Because I initially did not believe that a truly broad product could be assembled on deterministic systems. For local tasks — yes, they are perfectly suitable. Especially when you set up the flow yourself (not necessarily n8n) for the client. In fact, this is what we were doing in 2025.

What are non-deterministic AI systems? These are systems that can independently build a path to solve a task.

These are systems like Claude Code, where there are a bunch of specialized agents and orchestration that builds a solution based on the goal, context, and available information.

A deterministic system receives a task as input, processes it through a pre-built decision tree, and outputs a result.

A non-deterministic system receives a task as input and builds the decision tree itself. If something doesn't work out the first time, it can decide on its own whether another attempt is needed. And if agents are also given tools, context, and usage rules, that's where the magic begins.

We have now created a prototype agent for optimizing page texts: it corrects verbosity, removes AI clichés, and generally works on text quality. We provided it with optimization check tools, verbosity checks, connected a sub-agent for style correction, access to the database, and other context, specified what is connected to it and what it should do. And now the user can simply write what they roughly want, and the agent itself calls the tools, connects sub-agents, and builds a unique flow for the specific user.

And using this is much more pleasant than the usual workflow. Because in deterministic systems, for each type of task, you need to manually prescribe the entire scenario, conditions, branches, retries, and so on, but here you throw everything needed into one pile, and it all organizes itself.

And this is what I see as my goal for 2026 — to create something truly cool through non-deterministic agents.