Obsvra

Back to trackLesson 2 of 3

Agents 2: Not an AI Wrapper

“Agent-friendly” and “AI-powered” get used somewhat interchangeably, and when looking at a data tool, that can result in some healthy skepticism. A generator that’s actually an LLM asked to “make up some realistic claims data” means that the output isn’t really reproducible, not provably grounded in an actual statistical model, and depends on a live connection to whatever API is doing the generating.

If you’re curious or concerned about SpiceGrinder because it markets itself as “agent-friendly”, the important question is whether that agentic support is using a real, independent engine, or is just a wrapper around an LLM where the “engine” is a prompt.

That’s a claim you can check yourself, the same way Lesson 1 checked the Free/Pro tier claim — not by relying on the agent to tell you, but by looking at what actually runs.

Proof 1: Reproduce a value, consistently

Let’s revisit one of our canonical sample models: claim-frequency:

<dataset seed="42">
  <root node="ClaimObservation"/>
  <nodes>
    <Append name="ClaimObservation">
      <input name="Frequency"/>
      <input name="Severity"/>
    </Append>
    <Poisson name="Frequency" lambda="3.5"/>
    <Lognormal name="Severity" mu="8.0" sigma="1.2"/>
  </nodes>
</dataset>

Run it with seed 42 and the first row is always 5,1020.778556837366 — exactly, every time, on any machine. You don’t have to rely on a “plausible claim”; it’s Poisson(λ=3.5) and Lognormal(μ=8.0, σ=1.2) sampled from a seeded random number generator.

Ask an LLM-backed generator for the same thing twice and (outside of a statistical fluke) you’ll get two different, unreproducible answers, because there’s no seed to hand it and no fixed distribution underneath its output in the first place.

Proof 2: Take away the internet connection

Disconnect your network, or run SpiceGrinder inside a container with no egress at all, and generation doesn’t even notice. There’s nothing to notice — grind reads a model file, builds a generator tree in-process, and samples from it, with no outbound call anywhere in that path.

A tool whose generation is an LLM call can’t survive this test even in principle: without a network, there’s no API call it can make, and therefore no output to view. (A generator that depends on any other outside service call will also fail similarly.)

This isn’t a performance optimization or special operating mode SpiceGrinder happens to have — it’s a direct consequence of its design that makes SpiceGrinder completely self-contained, not needing an AI model (or any other external service) to generate data in the first place.

If your team has data residency requirements, or you’re running in an environment where an outbound call to a third-party AI API is a compliance or logistics problem on its own regardless of what’s in the payload, this is an actual answer, not just a policy promise.

Ok, so then how is SpiceGrinder “agent-friendly”?

Always optionally. You determine how much, if at all, you want to use agents.

Use our packaged skills or MCP server, or simply point to the documentation bundled in your install, and your agent will understand SpiceGrinder as a set of tools used to author and execute models.

You can ask your agent to help build a model: explain what you want it to do in natural language, and it will be able to suggest a starting point for a model that can then be refined by iterating with your agent or by hand.

For Pro users, point your agent to your model library and it should be able to include reusable pieces you or your teammates have already written, saving modeling time and increasing consistency between models.

Your agent will know how to run your model. Just tell it how you’d like to run it (locally via grind or point to a SpiceGrinder service you specify), how many rows you need, (optionally) what seed you’d like it to use, and where you want the results stored.

Use skills and hooks to have your agent run the model validator and analyzer tools as a routine part of the modeling process and on check-in to your revision control system. Build a test pipeline that creates test data from a model in your repository as part of your integration and regression test suites.

Because SpiceGrinder comes as a set of simple independent tools (invoked either as CLIs or as Java programs), you and your agent can arrange them in any way your workflow requires. Models are plain text, so they are easily created and understood by an LLM. The product documentation provides ample detail for agents to understand the system as a whole and how specifically to use it.

Summary: Verifying our claims

  • Reproducibility: same model, same seed, byte-identical output, every time — check it yourself by running the same file twice. Even better if you can run it in two different places.
  • No network dependency: turn off the network, watch it run exactly the same as it did with the network on.
  • Review the logs: grind --audit gives you the seed, the model, the node count, and a timestamp for every run — because what ran is a deterministic process that is recorded precisely.
  • The engine is code: every component’s behavior is a Java class that is running inside your environment. Nothing downloaded at runtime, no external API calls.

None of these require taking our word for it. You can verify it all yourself, and use SpiceGrinder Free to do it before you even consider upgrading to SpiceGrinder Pro.

If you ever find anything that indicates that the “agentic” part of SpiceGrinder gets embedded into the actual application suite — tell us and we’ll fix it. And then we’ll publicly explain where and how we got it wrong.