Kilo's data scientist built a business context layer to fix what AI agents were missing

AI agents have become far better at handling complex, code-adjacent work on their own, but that progress hasn't led to better business decisions, according to Pedro Heyerdahl, senior data scientist at Anaconda. Before Kilo Code became part of Anaconda, Heyerdahl was a one-person data team there, leaning heavily on AI agents to keep up. He presented his project at dbt Summit 2026.

His argument: technical capability was never the real gap. Pairing the dbt MCP, dbt skills, the Wizard CLI and dbt Semantic Layer with a well-documented dbt project lets an agent understand models, follow lineage and reason over metric definitions. But a dbt project is still a selective representation of the business, and without the surrounding organizational knowledge - the debates, failed attempts and open questions - an agent can get the numbers right and still misread the situation.

Heyerdahl tried common workarounds first. A single large memory file worked for a couple of weeks until the business moved on and the agent kept trusting stale information. Retrieval-augmented search over raw company material, and direct access to systems like Slack and Jira through MCPs, still left the agent deciding where to look and which source to trust when two disagreed.

So he built a pipeline, then rebuilt it twice. The first version summarized source material into Wikipedia-style markdown articles, which read well but lost detail each time the model chose what to keep. The second extracted atomic units of knowledge - facts, decisions, entities and prescriptions - preserving details such as who made a decision, when and why. The third added temporal validity and supersession, tracking when a fact was true and what replaced it, which he likened to a dbt snapshot for business context.

The current system routes work through an AI gateway, assigning cheaper models to routine extraction and frontier models to steps needing judgment, like drafting articles. A reflect step compares new information against existing facts and flags conflicts or likely-stale material for human review through a govern step, surfaced in a review interface and an MCP server. Agents can also write new findings back into the pipeline for human approval.

Source: dbt Labs Blog