With great context comes great autonomy

In a dbt Summit 2026 breakout session, Ben Moser, who works on the dbt platform team at dbt Labs, and Kevin Kim, product manager at Fivetran, made the case that the hardest part of building a trustworthy AI agent is everything it needs to know before it can be trusted to act, not the agent itself.

Context exists to build an accurate model of your business for an agent to operate in, the same way understanding what a stop sign means lets you drive safely without causing havoc. Without shared context, a simple question like "what was our ARR at the end of August" branches instantly into a dozen sub-questions: booked or current ARR, does it include services, which table is authoritative, who's allowed to see it.

"Context aims to create an accurate world for your agents to operate in," Moser said. "It helps them understand your business so that they can make the right decisions, answer your questions in the right way, and not cause havoc."

Moser framed context as a spectrum, not a single problem to solve all at once: simple semantic and structural context sits at one end, and far more complex situational or interpretive context, where meaning itself changes over time, sits at the other. Trying to boil the ocean and load everything into an agent's environment at once, he warned, is the wrong instinct.

A dbt project already builds a version of this map: models, tests, documentation, and lineage give agents a governed way to answer questions about anything that's been modeled. The interesting problems live at the edge of that map, in support tickets, Slack threads, and institutional knowledge that never made it into the warehouse.

Moser leaned into a mining metaphor, apt for a session held in Las Vegas: there's gold in that unstructured information, support tickets, sales calls, institutional knowledge, but leaving it unmodeled means agents either can't use it or use it unreliably. He framed the stakes in offensive and defensive terms. Used well, agents give a business a real edge; used carelessly, inaccuracies compound at machine speed in a way a single human error never could, burning both trust and token spend at the same time.

That's the gap Kim's team built Fivetran Context Layer to close. Connect knowledge sources like Notion and Confluence, operational sources like Salesforce, HubSpot, and Postgres, and analytics sources like dbt and Sigma, and Fivetran Context Layer parses all of it, builds a search index for semantic search, and infers an ontology that relates concepts across all three, so a revenue table in Postgres and a revenue doc in Notion get linked instead of living as separate facts.

Inside Agents Schema, a root table works like a glossary telling an agent what's in the schema at all, and a concept table stores company-specific shorthand, HRBP or SKO, for example, so an agent doesn't have to guess or search the open web for what your business means by an acronym.

Fivetran Context Layer writes what it learns directly into the warehouse as Agents Schema, a set of open, ordinary tables any AI client can query over MCP, rather than locking the context behind a proprietary API.

In a live demo, Kim showed an agent answering "which open opportunities have unresolved support tickets" by querying Agents Schema tables directly.

A new traces feature surfaces where an agent's confidence was low or context was missing, so analytics teams can curate and correct it over time.

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