Why AI Agents Are the New Consumers of Data with Tristan Handy (CEO @ dbt Labs)

Raised-bed gardening, then straight into LinkedIn algorithm gossip

The episode opens on a delightfully human note: Tristan Handy is excited about his raised-bed garden and the salad greens he just harvested, which he says tasted better than anything at Whole Foods. From there, Joe pivots into the real reason he called — Handy’s recent LinkedIn posts about agents — but not before the two go deep on Substack vs. LinkedIn and how discoverability now seems more topic-based, with saves and engagement depth mattering more than follower count.

The big thesis: agents, not humans, will be the biggest consumers of data

Handy says the industry is still focused on the wrong frame: conversational analytics where a human asks a bot a question. The bigger shift is that agents themselves will become the actors, querying systems on their own as part of business processes, which means data leaders need “agent analytics infrastructure,” not just prettier natural-language BI.

Why the current agent stack feels like old-school analytics all over again

One of Handy’s sharpest comparisons is to the 2015–2017 startup era, when companies would point BI tools directly at production Postgres and call it a day. He sees the same pattern now in the agent world — “just give it read access to Salesforce” — and says people are rediscovering, the hard way, that data pipelines solved real problems and didn’t exist just to annoy everyone.

Your warehouse bill may blow up before your token bill does

Handy’s practical warning is that agent loops use way more tools than humans ever could. In his own research workflows, agents perform vastly more web searches than he would manually, and he expects the same thing in business settings: agents will constantly look up data, hammer the lakehouse, and drive query volume “through the roof.” He predicts cost pressure will force teams to finally optimize, and he names DuckDB as a likely winner because it can handle a huge share of analytical queries far more efficiently than just brute-forcing everything in a big cloud platform.

Context stores, semantic layers, and the return of RAG-for-data

Handy argues that context stores are underrated not only for trust and answer quality, but for token efficiency. Too many data agents still do dumb things like stuffing an entire dbt manifest.json into the context window, which works until cost and accuracy collapse; what’s needed instead is a database-backed system that makes the right metadata easy to retrieve, with semantic layers as one component, not the whole story.

Stop building fake 'agentic data engineers' and fix the broken infrastructure

Joe says he gets pitched daily by startups claiming to build an “agentic data engineer,” and both agree the more interesting move is deeper: redesign the tooling itself. Handy gives a concrete example from dbt Labs’ own closed-loop agentic development — Git becomes a bottleneck, agents need separate worktrees and sandboxes, and their changes stomp on each other when merged — which is exactly the kind of infrastructure problem that appears when agents become first-class workers.

Agent-first organizations, Box’s layoffs, and the next startup model

The conversation zooms out into org design, sparked by Box CEO Aaron Levie’s essay about putting intelligence into the system rather than the humans and his 40% workforce reduction. Handy and Joe connect that to Conway’s Law, coordination costs, and the idea that future companies may be tiny, high-context teams with lots of agent leverage; Handy says if he started again today, he imagines something more like a six-person partnership spending heavily on tokens, not a classic venture-backed org chart racing to hire.

The internet flashback, the analyst future, and a plea to stay human

Both men get nostalgic about early internet life — 28.8 modems, public-library internet accounts, Mosaic, Napster, even printed phone books of websites — as a way to explain why AI still feels primitive but inevitable. Handy closes on a more grounded note: he’s using AI heavily at work but intentionally avoiding screens in personal life, and what excites him most is the chance to redefine the data analyst role so analysts spend less time recreating dashboards and chasing bugs, and more time shaping agent-powered business processes.