"The computers talk to us now" — Jason Ganz

Why this is in the vault

Jason zooms out from week-to-week agent benchmark noise and lands on a design point that directly informs MAC and the broader RDCO data-engineering thesis: code, docs, and configs written before ~2023 were authored for one audience (humans). Code, docs, and configs written today have to land for two audiences (humans + agents). He frames this through dbt internals that Drew Banin wrote in 2020 and that the team is now revising for an agent-readable world. The piece is essay-format, not curation - rare for AE Roundup, so it stands as a positioning artifact rather than a link aggregation.

The core argument

1. The category-shift is the obvious thing nobody is sitting with. "The computers talk to us now" is so true and so strange that it has stopped reading as strange. People answer agent demos with cost questions instead of awe.
2. Pre-2023 code was designed through a single lens. Drew's 2020 dbt internals were written for "the human analyst staring at a terminal at 4pm on a Tuesday." That was the only design lens that made sense at the time.
3. The audience for code has expanded mid-deployment. The same file, revisited five years later, now serves: human analysts, agents reading docstrings into context windows, coding tools generating dbt projects on the fly, and natural-language interfaces translating English ↔ configuration. The set of entities consuming the artifact grew without anyone updating the artifact.
4. Standard objections dispatched.
   • "SQL has always been talking to computers" → yes, but the conversation was one-sided; the new thing is general comprehension of intent, not a new vocabulary slot.
   • "It's just matrix multiplication" → true and uninteresting; the wing is still flying.
   • "Demos break in production" → partially correct, doesn't change the slope.
5. We are living inside continuous future shock and treating it as normal. Borrowing Toffler's term: when change exceeds the human bandwidth for digesting it, you don't feel awe or terror, you feel low-grade numbness. Jason's call is to stay rigorous AND keep feeling the weirdness.
6. Floor-under-feet predictions. Career trajectory of an AE entering in 2026 ≠ 2018. Shape of a data team in five years ≠ today. Documentation written two years ago is now driving measurable agent-efficiency gains.

Mapping against Ray Data Co

• MAC (TDD for data pipelines). Jason's framing strengthens the MAC thesis: if dbt internals from 2020 are now being revisited specifically because agents need machine-readable affordances, then the gap MAC is selling into - "your tests and contracts have to be legible to the agent that's writing your next model" - is a documented vendor pain inside dbt Labs itself. This is corroborating evidence for the bet, not a new insight.
• Dual-audience documentation as a vault discipline. The piece reframes a thing I already do (writing vault notes that are discoverable by future-me AND by an agent searching the graph) as a generalizable design principle. Worth surfacing in Sanity Check at some point: "every doc you write now has two readers, and most teams haven't internalized the second one."
• Analytics Engineering positioning. This is dbt Labs' editorial voice planting a flag on "we are the company designing for the agent-readable world." That matters for who MAC partners with downstream and whose ecosystem gravity to ride. Tristan's "BI's Second Unbundling" piece (2026-05-03) sat in the same editorial direction; this is the human-stack-side companion.
• Sanity Check angle (not derivative). A derivative piece would restate Jason's framing. The non-derivative angle: founders/operators are also dual-audience now - your Notion docs, your job descriptions, your investor updates are being read by agents working on behalf of other humans. That's not the AE Roundup story. That's an RDCO story.

Editorial note

This is the second consecutive AE Roundup issue (after "BI's Second Unbundling") that drops the curation format in favor of a single-author thought piece. Watch whether this is a permanent format shift or two one-offs. If it's a shift, the deep-fetch budget on AE Roundup becomes irrelevant going forward and the vault treatment changes from "curation extract" to "thought-leadership single-source."

Related

• [[06-reference/2026-05-03-ae-roundup-bi-second-unbundling]] - prior AE-Roundup, same shift to thought-leadership format, Tristan on the BI side
• [[06-reference/2026-04-29-data-engineering-central-ai-changing-de-fast]] - parallel "AI is changing data engineering fast" framing from a different curator
• [[06-reference/2026-05-09-data-engineering-central-cognitive-overload-ai-development]] - the cognitive-load counterweight to Jason's "feel the awe" framing
• [[06-reference/2026-05-04-dataengineeringweekly-268-agents-replacing-search]] - agents-as-primary-interface evidence from another data-engineering venue