The rise of the individual contributor

Between December 2025 and May 2026, the agent-hours the system absorbed per hour of the operator’s attention rose roughly fivefold (Exhibit 1), because the job of Tristan — the operator behind Data Driven Partners — changed from typing, one conversation and one correction at a time, to specifying work precise enough to run without him. The evidence follows: a month-by-month record, then the record’s hours cut three ways — when they happened (Exhibit 3), whose they were (Exhibit 4), and where they pooled (Exhibit 7).

Across four tools — Cursor, Claude Code, Codex, and Claude desktop — the full record runs to roughly 1,350 system-active hours of data work in 802 sessions, steered by 12,565 of his own messages: specification, correction, review. What the transcripts document is the assembly, failure, and reassembly of an end-to-end agentic data pipeline — sourcing, enrichment, identity, certification, sale, self-monitoring — that today delivers records at about $1 a lead with 24-hour turnaround, re-verified on demand rather than left to decay in a static catalog.

He runs a cold-email agency. It runs on business-contact data — who works where, what their address is, whether it still resolves — and that data decays at two to three percent a month, so a third of any rented list is wrong within a year (see “The economics” comparison table). In November he decided to stop renting it and build his own. The first move was a migration: drag the data off SQL Server, where the Azure bill was climbing, onto Postgres — the opening experiments in the store lane of Exhibit 2.

He wired an assembly line: take a company name, find its real domain, find the decision-maker, find the email, verify it, enrich it. Each step a worker; each worker its own small machine. Most of them broke. Jobs fell into retry loops that claimed a task, crashed before finishing, and handed it back to themselves. A second failure class was harder to detect: functions returned “200 OK” while the real work died in the background, so the dashboards reported green over a stalled queue.

December took 304 hours of work, nearly all of it one person in the Cursor editor, one message at a time — the tall single-tool bar at the left of Exhibit 4. The longest conversation in the record belongs to this regime: it opens on December 19 with a request to implement a pipeline design in Supabase and closes a month later, more than a thousand messages deep, with a working domain enricher and a store holding 37.2 million prospects. The instruction that repeats most in the logs is not a request for anything new. It is simple, simple, simple — tear the thing down and rebuild it with fewer moving parts.

He started re-deriving the discipline he’d never been taught. He had an assistant take on the methods of Ralph Kimball, the patriarch of data-warehouse design, and audit his work; he collapsed a tangle of tables into one jobs table where all the progression logic could be read in one place. Where a rule would do the job of a model, he used the rule: deterministic filters dropped foreign domains, broken characters, and anyone who wasn’t a founder or an owner before a cent was spent on them.

Then identity. The same company kept entering the store under three spellings, and the duplicates degraded the records that actually sold. The fix removed a problem rather than adding intelligence: rather than build a smarter merge, he made identity permanent and attributes perishable — mint an internal ID once, let domains and emails decay against it, and resolve duplicates on deterministic match keys (Exhibit 5, “Company identity”).

The next constraint was sourcing. Renting records made no sense for the supply side either, and cloud scraping failed on cost and on static-IP limits. So he moved from renting compute to owning it: a domain-finder scale-up and hand-written scrapers — a 190-hour workstream, the second-largest line in Exhibit 7 — then an Ubuntu laptop reachable over Tailscale, set up as a dedicated scrape box that fed the pipeline its raw material (Exhibit 2, “Sourcing & scraping”).

The change behind the fivefold figure came in the spring (Exhibit 1). Through the winter he had worked one conversation at a time, supervising every step; one logged session ran 186 hours. In April he stood up the scaffolding — specifications precise enough to run unattended, a load balancer to pool model capacity — and in May he ran fleets of agents in parallel. In May the system absorbed 524 hours of agent work, most of it running on its own while he steered: the dense right-hand cluster in Exhibit 3, and the month in Exhibit 4 when Codex and Claude Code swarms displaced the editor. The architecture caught up with the work: an append-only observation log, a trust-ranked golden record, a certified-leads view, and a freshness policy — build the data product, not just the process that fills it.

By June the system largely operates itself. Scheduled monitors check the morning’s pulls, a daily job chases anything stuck, and the system refuses to sell a record it can’t vouch for; the seven capability rows of Exhibit 5 summarize the November-to-June difference. He reads the summary. The next problem is already measured: reply rates on his outbound have fallen from twenty qualified leads a day to three — the number he is pointing the machine at next.