Summary

Simon Späti argues that AI agents are most valuable in data engineering when built on a deterministic correctness layer, not used as general-purpose chat tools. He outlines three levels of AI agent maturity — chat-phase, autonomous CLI-access, and dedicated tooling — with dedicated tooling being the most reliable because it offloads factual operations to deterministic engines. The core risk he identifies is not errors that crash pipelines, but silent wrong numbers: queries that compile and pass tests while introducing incorrect joins or inflated metrics. He uses Altimate Code as a concrete example of the architecture he advocates: a probabilistic LLM layer on top of a deterministic Rust/TypeScript core that handles SQL parsing, validation, and equivalence checks sub-millisecond. The post includes a practical blast-radius analysis example showing how such tooling can automatically map downstream impact before a column rename. His conclusion is that general-purpose agents are fine for exploration, but any output touching production requires a deterministic core underneath.

Key Insight

AI agents become trustworthy in data engineering only when the probabilistic LLM layer is separated from factual operations — SQL parsing, validation, lineage — which must run on a deterministic core that proves correctness rather than guessing it.

Spicy Quotes (click to share)

• 7

  The danger isn't only a crash or an error message, but a wrong number that didn't break.

• 8

  Bare agent use might be cheap, but only until they're wrong, and then the cost is unbounded.

• 5

  Like a compiler, the agent never decides whether two queries are equivalent or a column exists upstream.

• 4

  The key is to get use out of AI, not to get more work.

• 6

  Building from scratch with no knowledge or seniority is dangerous. Why? Because they can't verify if the produced code is correct.

• 6

  Cost-per-token is the wrong number to optimize, since the meter itself can move with a vendor's next model update.

• 5

  The model matters less, but the structure does.

• 4

  Not every task needs a level-three agent. A quick chat-phase agent is fine for exploring a dataset or drafting a query you'll review yourself. But the moment that output touches production, you want the deterministic core underneath it.

Tone

practical and opinionated, with a sponsored/advocacy lean toward Altimate Code