“Five things I believe about the future of analytics” — Tristan Handy

Why this is in the vault

Handy explicitly names the harness layer as a leverage point for analytic agents and predicts agent-initiated queries will outnumber human-initiated queries 100x within 36 months. This is the dbt creator publicly endorsing two RDCO load-bearing theses (harness engineering and agent-as-primary-data-consumer) in the same essay, with a citation to a fresh Stanford paper (Meta-Harness) that the vault did not yet have. Strong signal for the harness-thesis cluster and the consulting positioning narrative.

The core argument

Five interlocking claims that ladder to one conclusion: design data infrastructure today for an audience that is mostly agents, not humans.

1. Analysts are going technical. Vibe coding pulls finance and data analysts onto the command line — out of BI tools, into Claude Code, Cursor, IDEs. Data consumers stay in BI/Notion/Claude Desktop. Bifurcation: everyone except consumers moves into technical tooling.
2. Data usage will explode. The pipeline layer got 10x better over the last decade; the analysis layer didn’t, because analysis is “thinking” and we hadn’t solved thinking. With LLMs, that bottleneck vanishes. Disruption at the usage layer dwarfs disruption at the pipeline layer.
3. Analytic agents are happening now. dbt Labs internal, Meta (weekend prototype to thousands of users in six months), OpenAI’s analytics agent, Ramp Research. Production today, working, being adopted because they work — not experiments anymore.
4. Agents consume dramatically more than humans. The new query class is agent-initiated — agents generate the hypothesis, write the query, follow chains of reasoning, decide what’s worth surfacing. Only the final “interesting enough to tell someone” result reaches a human. dbt MCP server calls grew 50% MoM every month since launch. Handy: agent-initiated queries will surpass human-initiated within 12 months at most companies; 100x more agent than human within 36 months is plausible-to-conservative.
5. Harnesses are a leverage point. Cites Stanford’s Meta-Harness: End-to-End Optimization of Model Harnesses — the harness around a fixed LLM can produce a 6x performance gap on the same benchmark. Vertical-specific harnesses (tuned to a domain) significantly outperform generic ones, and harness search can be automated.

Conclusions Handy explicitly draws:

• Agents become the primary consumers of analytic data within 12 months — design infrastructure for that now.
• Analyst impact goes up; the job changes. Survivors build and operate agentic analytic systems, not dashboards.
• Human-facing analytic assets still get built — but inside Claude Code/Cursor, not BI tools, because the creators have moved.
• Analyst workflow starts to resemble a front-end software engineer’s — a UI layer on top of a data-driven reasoning engine.

Mapping against Ray Data Co

• Harness-thesis reinforcement (high signal). Thing 5 is a near-verbatim restatement of the thin-harness-fat-skills argument, now from the dbt creator citing fresh Stanford research. Adds an authority anchor for the cluster: Tan, Pachaar, Greyling, Chase, and now Handy + Stanford. Worth pulling Meta-Harness itself onto the research backlog — the 6x performance gap claim is the kind of citation that lifts a Sanity Check piece on harness engineering from “asserted” to “evidenced.”
• Agent-initiated query thesis is the consulting headline. “Agents will be the primary consumers within 12 months — design for that now” is a sellable urgency frame for prospects still building dashboards. Pair with the dbt MCP 50% MoM growth as the proof point. This is stronger than the generic “agents are coming” pitch we’ve been hedging into.
• Vertical-harness finding validates RDCO scope discipline. “Vertical-specific harnesses outperform generic ones” is the academic case for building data-domain harnesses (the vault, the skills, the dbt knowledge) rather than chasing horizontal frameworks. RDCO’s bet is exactly a vertical harness for the founder-COO + data-consulting domain.
• Bifurcation framing is a content angle. “Everyone except consumers moves into technical tooling” is a clean three-line opening for a Sanity Check on tool stack rationalization. Map cleanly to Levie’s agent-deployer JD piece and the dbt MCP/text-to-SQL benchmark.
• Counter-argument tension to flag. Natkins (Semi-Structured) argues data layer, not harness, is the moat. Handy here puts harness front and center while still implicitly relying on dbt’s semantic+model layer as the substrate. The honest synthesis: harness and data layer are both moats, neither alone — which is what the harness-thesis-dissent synthesis already concluded. Handy’s piece does not collapse that tension; it leans harness.

Related

• 2026-04-12-ae-roundup-move-up-the-stack — Handy’s “change management is the real barrier” piece; this issue is the technical/architectural complement
• 2026-04-05-ae-roundup-moving-up-the-stack — Jason Ganz’s earlier framing of the same phase change; Handy now cites adoption data (50% MoM) backing the prediction
• 2026-03-31-semistructured-data-layer-does-the-work — Natkins’ counter-frame: data layer, not harness, is the moat. Productive tension with Thing 5 here
• synthesis-harness-thesis-dissent-2026-04-12 — the existing dissent synthesis; Handy’s piece adds an authority data point on the pro-harness side and a fresh academic citation (Meta-Harness)
• 2026-04-08-better-harness-evals-hill-climbing — the systematic methodology for the harness optimization Handy says Meta-Harness automates
• 2026-04-07-dbt-semantic-layer-vs-text-to-sql-benchmark — the empirical case for the semantic-layer substrate Handy’s agent-initiated queries depend on
• 2026-04-11-garry-tan-thin-harness-fat-skills — the originating articulation of the thesis Handy is now extending into the analytics domain
• 2026-04-12-alphasignal-claude-code-leak-harness-engineering — independent evidence that frontier labs are investing in harness engineering

Tracked-author candidates / follow-ups

• Meta-Harness paper (Stanford) — file as a research-backlog item; the 6x performance gap citation is load-bearing for any future harness-engineering Sanity Check.
• Meta’s internal analytics agent writeup (linked from the issue) — worth a separate vault note as a case study of weekend-prototype-to-thousands-of-users-in-six-months adoption curves.
• Ramp Research — already on RDCO radar; Handy’s repeat citation reinforces the case study.