The stack
Four layers, top to bottom
Agentic analytics is usually discussed as though it were a model choice. It is four layers, built by different people, failing in different ways, replaceable on different timescales. Being able to name which one is at fault is the difference between fixing a system and swapping models and hoping.
The agent, in three parts
A model reasons. A harness runs the loop, decides what the model sees, and handles failure. Interfaces carry the request out to the world. Collapse the three and you have bought a product rather than built a capability.
- 04 The agentic stack Three separable layers, model, harness, and open interfaces, each with its own failure mode. Most agent problems are misattributed across them.
- 05 The model What the reasoning layer actually contributes, what it cannot know, and why treating model choice as the whole architecture leads people astray.
- 06 The harness The program that runs the loop. It decides context, tools, retries, memory, and when to stop, which is where most agent behaviour actually comes from.
- 07 Open interfaces The four contracts an analytics agent speaks through, why each should be a specification rather than a product, and what a closed one costs later.
Underneath all three
Governed semantic data
The layer that decides whether the answer is right. One definition of each metric, one copy of each dataset, one place access is enforced, and enough history to reproduce an answer after the fact.
Five Apache projects settle five separate questions here, and the reason to care that each is an open specification is that agents multiply readers. Formats one engine can read cannot serve many readers.
The substrate, entry by entry
What each project settles, where it stops, and why the boundary between them is deliberate.
- Why an open lakehouse Agents multiply the number of readers and the number of questions. That pressure breaks architectures that assume one engine, one copy, and predictable queries.
- The semantic layer as an agent contract The place where a business writes down what its terms mean. Without it, an agent guesses at definitions, and a fluent guess is worse than an error.
- Apache Iceberg The table format that lets many readers and writers share one set of files safely, which is exactly the condition agentic analytics creates.
- Apache Polaris An open catalog that resolves table names, decides who may read them, and hands out credentials scoped to just those files. The enforcement point agents cannot route around.
- Apache Arrow A standard in-memory layout for tabular data, so results move between the engine, the tool, and the agent without being repeatedly converted.
- Apache Parquet The columnar file format most lakehouse data sits in. Its statistics are what let an agent's exploratory query read a fraction of the table instead of all of it.
- Apache Ossie An incubating project defining a vendor-neutral standard for semantic metadata, so a metric definition is portable rather than trapped in one tool.