Open Data Infrastructure in the age of AI agents
Why this is in the vault
This is a reunion panel — George Fraser (Fivetran CEO), Tristan Handy (dbt Labs CEO), and Martin Casado (a16z GP, board member at both companies) are the same three voices already filed in [[2026-04-03-great-data-debate-a16z]], a 2020 a16z podcast on data-lake-vs-warehouse convergence — now revisiting the open-vs-walled-garden question specifically for AI agents, and landing on an architecture argument (own a modeled copy of your data; don't wire agents straight to SaaS MCP servers) that doubles as validation for how RDCO's own vault + knowledge-graph setup is built, and as a sharper pitch angle for the phData credibility project.
Mapping against Ray Data Co
The load-bearing connection is architectural, not just topical: George Fraser's prescription — "you want a copy of all of your business's data that you control," modeled and in an open format (Iceberg, in his example), so an agent queries a governed layer instead of raw systems of record — is the same split RDCO's own setup embodies, imperfectly. The vault + qmd search + graph.duckdb knowledge graph is the "modeled, controlled copy" half of that architecture (deliberately structured, versioned, queryable); but a chunk of daily agent work still runs the pattern Martin Casado explicitly warns against — "MCP straight to every system of record" (Gmail, Notion, Slack MCP calls fired live, no modeled layer in between). Casado's point that MCP is "like HTTP: a protocol that says nothing about structured data, guarantees, or security," and Tristan Handy's add that "permissions get tangled once more than one system or one person is involved, and tokens get expensive fast," is a concrete argument for routing more of RDCO's own MCP surface through the modeled vault/graph layer rather than ad hoc live calls — worth a deliberate look rather than a reflexive one, since Ray's agent is single-operator (no multi-person permission tangle yet) and token cost on these calls is not currently a measured constraint. Second, sharper mapping: Tristan Handy's framing that unlocking agentic analytics is "80% of the work you've already been doing" (centralize data, model it well, add metadata) plus thin plumbing on top is close to a ready-made pitch angle for [[project_credibility_for_phdata_sales]] — it reframes phData's actual deliverables (dbt modeling, warehouse architecture) as the load-bearing work behind "agents in production," not a separate AI initiative bolted on after.
The core argument
Companies are building their own AI data analysts (Stripe, Ramp, DoorDash, Fivetran's internal "Context Layer") instead of buying a packaged tool, because the bottleneck was never analyst supply — agents provide an effectively infinite number of them — it was the long tail of tedious, ambiguous questions no human team had time to chase. Ramp's internal research agent, launched in Slack, went from an expected 2x bump in question volume to 20-30x within weeks and 50x within six months: "that's the latent demand in the organization to know things... it had been too annoying to get the answers, so people got by without them." George Fraser's own example: a casual guess at two Fivetran product-penetration metrics was off by roughly 2x and 4x against the real numbers his Context Layer agent pulled in seconds — "when you don't know the exact answer, the vibes in the air are often very wrong." Martin Casado frames the shift as productivity, not headcount reduction, across the companies he has visibility into as an a16z board member: "you only do headcount reduction if you don't know how to grow... if you know how to grow, you need the humans behind it."
The architecture argument follows from that: in a transformation where "nobody knows what the stack is going to be" (Casado's example — Cursor went from tab-complete to IDE to agent to foundation-model company within two years), the market rewards the open, composable stack over the walled garden. George Fraser's concrete recommendation is to hold a controlled copy of business data in an open table format (Iceberg, "the one everyone supports"), so whatever tooling emerges next can read it without renegotiating access. The competing, rejected architecture is wiring agents straight into Salesforce/Workday/Jira over MCP: Casado insists MCP is an access protocol with "nothing to do with structured data, guarantees, or security" — comparable to HTTP, not to a data architecture — and Handy adds that permissions and token costs both break down once more than one system or person is involved. The episode also flags vendor incentive friction directly: SaaS vendors increasingly claim ownership of "their" data and are testing API tolls (Amazon's short-lived attempt to charge sellers to view their own sales-volume data via the Selling Partner API is the named example), which Casado reads as a predictable, temporary resistance pattern rather than a permanent one.
⚠️ Sponsorship
This issue carries the standard Analytics Engineering Roundup "sponsored by dbt Labs" line and an inline house-promo block for dbt State (a new dbt feature, 30-day trial) — self-promotion by the publisher, flagged per the sender's standing pattern, not a third-party paid placement.
The more consequential bias is structural, not a disclosed sponsor line: all three speakers have a direct commercial stake in the conclusion the episode reaches. George Fraser is Fivetran's CEO; Tristan Handy is dbt Labs' CEO; Martin Casado is an a16z general partner who has personally invested in both dbt Labs and Fivetran and sits on both companies' boards (stated directly in the newsletter's framing of the conversation). The argument — keep your stack open, own a modeled copy of your data, don't let SaaS vendors wall it off, don't let MCP-to-everything replace a real data architecture — is also exactly the argument that keeps Fivetran (data movement into a controlled store) and dbt (the modeling layer on top of it) commercially relevant in an agentic-AI world where a plausible alternative pitch is "skip the pipeline vendors, point the agent at the SaaS tools directly." None of the specific facts in the episode (Ramp's question-volume growth, Fivetran's own metric-accuracy example, the Amazon API-toll walk-back) look fabricated or unsupported, but the overall framing is three people with aligned financial interests converging on the architecture that preserves those interests — read the "don't connect agents directly to SaaS over MCP" argument as a vendor position as much as a neutral technical one.
Related
- [[2026-04-03-great-data-debate-a16z]] — the same three speakers (Fraser, Handy, Casado) in a 2020 a16z podcast on data-lake-vs-warehouse convergence; this episode is a direct thematic sequel applied to AI agents
- [[2026-09-20-dataengineeringweekly-288-agentic-data-stack-dbt-summit]] — same dbt Summit moment, same "version-controlled, modeled layer between agent and raw data" argument from a different sender
- [[2026-08-13-analytics-engineering-roundup-lowin-mcp-skills-bazooka]] — same sender's prior single-guest podcast issue, also on MCP's limits as a sole agent-to-system interface
- [[2026-10-09-analytics-engineering-roundup-personal-agents-jev-decision-models]] — same sender's immediately preceding issue
- [[project_credibility_for_phdata_sales]] — the phData pitch angle this episode's "80% of agentic analytics is work you've already done" argument sharpens
- [[project_channels_agent_setup]] — RDCO's own Mac-mini agent, whose vault+qmd+graph.duckdb layer is the "modeled, controlled copy of your data" half of the architecture this episode argues for, against the live-MCP-to-SaaS pattern it also partly runs