How a data scientist at Kilo built the business context layer his AI agents were missing

Pedro Heyerdahl, senior data scientist at Anaconda, was a one-person data team at Kilo Code before it became part of Anaconda. He relied heavily on AI agents to keep up, and introduced us to his project at dbt Summit 2026. His core argument in this session: agents have gotten much better at working through complex, code-adjacent problems on their own, but that improvement hasn't translated into better business decisions, because technical capability was never the actual gap.

Heyerdahl's premise is that combining the dbt MCP, dbt skills, the Wizard CLI, and dbt Semantic Layer with a well-documented dbt project gives an agent the ability to understand models, follow lineage, and reason over metric definitions. But a dbt project and its data are still a structured, selective representation of the business: what a company decided to track, how it agreed to measure it, and the decisions, failed attempts, and open questions that shaped it along the way. Without that surrounding context, an agent can get the numbers right and still misread the situation.

He framed the real gap as organizational knowledge, the kind that lives in Slack threads, meeting debates, and pull request explanations, none of which were created to be a durable record.

He described trying several common workarounds before landing on a different approach. Dumping every available file into one large memory file worked well 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 giving agents direct access to systems like Slack and Jira through MCPs, both still left the agent to decide where to look and which source to trust when two disagreed.

So he built a pipeline, and rebuilt it twice more. The first version summarized raw source material into Wikipedia-style markdown articles, which read well but quietly lost detail every time the model chose what to keep. The second version stopped summarizing and instead extracted atomic units of knowledge: facts, decisions, entities, and prescriptions, each preserving different information such as who made a decision, when, and why.

The third version added temporal validity and supersession, tracking when a fact was true and what replaced it, which he compared to a dbt snapshot for business context, so a superseded decision could still be retrieved for historical analysis without polluting current search results.

The current system routes work through an AI gateway that assigns cheaper models to routine extraction and frontier models to steps that need 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, closing the loop.

Check out Fivetran Context Layer if you want to turn your data sources into the context agents need to understand your business and give answers you can trust.

Watch every keynote and breakout from dbt Summit 2026 on demand. Watch dbt Summit 2026 on demand →