06-reference

every thesis 2027 recap

2026-11-08·reference·source: Every·by Grant Harvey
everycompound-engineeringharness-engineeringagent-oversightevalsconference-recap

What You May Have Missed at Thesis: 2027

Every's recap of its first in-person conference, Thesis: 2027 (Pioneer Works, Brooklyn, ~400 attendees + livestream), organized around Dan Shipper's framing question: "What does great human work look like after automation?" Byline: Grant Harvey (Every staff writer), with editorial support from Eleanor Wikstrom; quotes Kate Lee (EIC).

Why this is in the vault

A multi-session conference digest that happens to concentrate two load-bearing RDCO concepts — agent-oversight-at-scale and evals-as-reusable-checks — into concrete, named examples (Notion's kanban-for-agents, Every's own "Vibe Checks to Evals" workshop) rather than abstractions.

Mapping against Ray Data Co

The single most concrete connection: Notion cofounder/CEO Ivan Zhao's warning, relayed from his interview with Dan Shipper — "I don't think we should surrender our understanding" — paired with the mechanism he describes: when every engineer on a team runs their own coding agent, a medium-size codebase can stop making sense to the people responsible for it within a week. Notion's fix is a shared kanban board that lets humans see what the agents are doing. This is a named, operating instance of exactly the problem RDCO's harness-engineering thesis exists to solve (see [[2026-07-26-harness-seven-failure-mode-scorecard]] and the Thariq 1M-context note already anchored in AGENTS.md) — and it's independent validation that a large, agent-fluent org (Notion) converged on "visibility layer over agent work" as the fix, not more autonomy. Worth testing whether RDCO's own station-brigade (spec→test→code→critic) would read as legible on a kanban board the way Notion's does, or whether it's currently a black box between dispatch and handback.

Two of the "Sessions You Might Have Missed" picks map directly onto RDCO's own tooling, by name, from inside RDCO's own publisher (Every): Mike Taylor's "From Vibe Checks to Evals" — "turn every correction you make into a yes-or-no check you can rerun on each new model... you don't have to be very good at evals, you just have to have an opinion" — is a plain-language restatement of what RDCO's verify-* stack already does structurally (see [[02-sops/2026-05-20-verify-stack-two-gate-pass-fail-architecture]]: LLM-as-judge gate + deterministic code-eval gate). And Kieran Klaassen/Trevin Chow's "Where Humans Belong in Compound Engineering" (the "compound sandwich," what to automate based on stakes) is the same Compound Engineering lineage RDCO has tracked across multiple prior Every pieces ([[2026-04-04-compound-engineering]], [[2026-07-13-every-polish-agent-built-software]]) — this issue adds "automate based on stakes" as the explicit decision rule, which is a cleaner cut than RDCO's current ad hoc judgment calls on what gets a critic gate vs. runs unreviewed.

No new RDCO-specific action falls out of this issue on its own — it's corroboration and vocabulary, not a new finding — but it's a useful external checkpoint that the harness-oversight and evals-as-cheap-checks framing RDCO has been building toward is where the frontier operators (Notion, Every itself) actually landed in practice, in front of an audience of enterprise AI-adoption leads (Pfizer, Royal Caribbean) who are RDCO's own target buyer profile for the "agents in production" credibility thesis ([[project_credibility_for_phdata_sales]]).

No third-party deep-fetches were needed: the email body itself carried enough concrete, attributable detail (named sessions, direct quotes, specific mechanisms) to assess the mapping without pulling the linked session pages or Every TV videos.

Curation section

Notes from Thesis: 2027 (keynote + panel recap)

Sessions you might have missed (Every's own picks)

What people said

Attendee quotes ranged from "Masterclass conference" (Andrew Ambrosino, OpenAI) to best-dressed-crowd praise — standard conference social proof, flagged here only because the attendee list itself (Pfizer, Royal Caribbean, U.S. Bank) is evidence of who's buying enterprise-AI-adoption thinking right now.

Related