"Compound Engineering Gets an Upgrade" — Kieran Klaassen

Why this is in the vault

Klaassen (who coined "compound engineering" ~a year ago) is updating the canonical framework: the original four-step loop has grown to eight. The shift is not cosmetic. His claim is that as models got better, the work phase became "boring in the best way" — agents reliably plan, code, and test — so the human's leverage migrated to the two ends of the process: deciding what is worth building, and judging whether the finished thing actually feels right. He generalizes this beyond code to knowledge work: the middle gets automated; quality lives at the bookends. Directly relevant to RDCO because compound engineering is a sibling of RDCO's own multi-agent build pipeline and verification discipline, and this revision is essentially a real-world admission that the same two gaps RDCO already instruments (front-end intent, back-end taste) are where the residual human value sits.

The core argument

• Original loop: brainstorm → work → review → compound → repeat. Still the engineering core.
• New loop: ideate → brainstorm → plan → work → review → polish → compound → repeat.
• Two additions: ideate/brainstorm at the front (what's worth building, the user, the edge cases, what's exciting enough to spend time on) and polish at the end (click around, read the copy, ask whether it feels right even when it technically passes review).
• The metaphor (credited to collaborator Trevin Chow): the "AI sandwich" — AI is the filling, humans are the bread on both ends.
• Compound remains the load-bearing step: each feature should make the next one easier.
• Generalization: the pattern applies to knowledge work broadly, not just engineering.

Mapping against Ray Data Co

Strong overlap. RDCO already runs a more formalized version of this loop, which makes the comparison a useful audit of gaps rather than a new idea to adopt.

Where RDCO already does this (often more rigorously):

• Work / review / compound map cleanly onto RDCO's 12-stage production-mode workflow and the multi-agent pipeline seats (spec-author → test-author → code-author → critic; see [[2026-05-12-multi-agent-pipeline-config-schema]]). RDCO's "review" is harder than Klaassen's — it routes through a fresh-eyes independent worker ([[2026-05-19-verification-as-independent-worker-pattern]]) precisely because the builder shows confirmation bias on its own artifact. Klaassen's review is still mostly self-review.
• Compound is RDCO's strongest match: feedback memories, SOPs, and the skillify loop are the codified "each unit makes the next easier" mechanism. RDCO arguably out-instruments Every here — memory files persist across sessions, not just within a guide.
• Polish maps to RDCO's design-critic / video-critic / verify-pdf-output gates and the "no slop cannon" production-mode rule (IC-mode vs production-mode). RDCO already separates "technically passes" from "feels right" via dedicated taste-scoring critic agents.

Where the framing adds something / exposes a gap:

• Ideate is the genuinely useful add for RDCO. The pipeline is strong from spec onward, but the pre-spec "is this even worth building" step is comparatively under-instrumented. RDCO has the targeting-system prioritization filter (targeting / instrumentation / tools / feedback loop) which is the analog — but it's a judgment heuristic applied ad hoc, not a named first stage of every build. Klaassen's framework argues ideate deserves to be a first-class step, not a vibe.
• The V-model discipline RDCO uses (spec/test pairing before build) is absent from Klaassen's linear loop. RDCO's pipeline is arguably ahead here: test-author runs independent of code-author, which compound engineering's flat 8-step sequence doesn't encode.
• The honest admission that "work is now boring" matches RDCO's lived experience that reversible execution should just run (auto-mode signal-to-noise) and human attention should be reserved for decisions/surprises/blockers — i.e., the bread, not the filling.

Net: medium-to-strong. No new mechanism to adopt, but "ideate as a named first stage" is a small, real upgrade to how RDCO frames the front of its pipeline, and the note is good corroboration that RDCO's verification-as-independent-worker discipline is ahead of the public state of the art.

Provenance note

Bylined essay, plaintext rendered fully via Gmail (from-source, not reconstructed from web). Promotional content present but not paid sponsorship: a CTA for Every's "Compound Engineering Camp 3" (June 5, paid subscribers) and a standard footer plugging Every's own products (Spiral, Sparkle, Cora, Monologue, Proof). Author is GM of Cora, an Every product — so this is house thought-leadership with an inherent self-interest in promoting the Every AI-tooling ecosystem. No third-party sponsor.

Related

• [[2026-05-01-trevin-compound-engineering-v3-4]]
• [[2026-01-30-every-compound-engineering-framework]]
• [[2026-02-09-every-compound-engineering-guide]]
• [[2026-04-22-every-bread-in-ai-sandwich]]
• [[02-sops/2026-05-19-verification-as-independent-worker-pattern]]
• [[02-sops/2026-05-12-multi-agent-pipeline-config-schema]]
• [[2026-05-28-every-vibe-check-opus-4-8]]