Compound Engineering

Each unit of engineering work should make the next one easier, not harder: that is the canonical formulation Kieran Klaassen and Dan Shipper set out for Compound Engineering at Every, where agents write 100% of the code and five products (Cora, Spiral, Sparkle, Monologue, Proof) are each built and maintained mostly by one person. Traditional development injects complexity with every feature; the claim here is the opposite slope, where bug fixes eliminate whole categories of future bugs.

The operational core is a loop. The four-step version splits time roughly 40% Plan, 10% Work, 40% Assess, 10% Compound, with 12 parallel subagents reviewing security, performance, complexity and architecture. The February 2026 guide expands it to seven steps (Ideate, Brainstorm, Plan, Work, Review, Polish, Compound) and adds the 50/50 rule: half the effort on features, half on improving the system. Claude Code is the main tool, though Shipper and Klaassen describe the approach as tool-agnostic.

By v2.60.0 (31 March 2026), Trevin Chow had made code review mandatory across the pipeline, paired with a six-level confidence rubric and a 0.60 suppression threshold that cut false positives by 49%. Klaassen's Thinkroom guide locates the binding constraint in retrieval: thirty-five learnings sit in docs/solutions/ with frontmatter built entirely for search, because postmortems otherwise rot in wikis.

Shubham Saboo reaches a similar shape from product management with Loop Engineering, and Stack Overflow for Agents from public knowledge exchange. Staleness stays unresolved: if knowledge is what compounds, a stale learning compounds too, which is why Klaassen adds read-time, write-time and on-demand refresh checks.