AI EngineerVideoSource Linked

The Era of Compound Engineering — Kieran Klaassen, Every/Cora

Compound engineering treats each agent mistake and product judgment as reusable context. The useful discipline is to invest in memory and review so the next feature becomes easier, not merely faster.

AI Engineer · Aug 20, 2026

Open Source Open Markdown Open JSON

Source Summary

Kieran Klaassen describes a solo loop for shipping Cora: brainstorm, plan, execute, review, polish, and compound the learning. His operating rule assigns **50% of the effort** to delivering the feature and **50%** to teaching the system what should improve next time.

Practical Implication

Capture decisions, postmortems, product taste, and repeated corrections in agent-readable files or skills. Keep human judgment at the start and end of the loop, while letting agents handle implementation only after planning, tests, and review are dependable.

Agent-Ready Context

Kieran Klaassen describes a solo loop for shipping Cora: brainstorm, plan, execute, review, polish, and compound the learning. His operating rule assigns **50% of the effort** to delivering the feature and **50%** to teaching the system what should improve next time.

Capture decisions, postmortems, product taste, and repeated corrections in agent-readable files or skills. Keep human judgment at the start and end of the loop, while letting agents handle implementation only after planning, tests, and review are dependable.

The claim that implementation is mostly solved comes from one builder's workflow, not comparative evidence. Long autonomous runs still depend on strong evaluation, and accumulated memory can become noise unless lessons remain specific, current, and retrievable.

Connected Context · Feed7 Judgment

This provides the operating philosophy behind compound-engineering toolkits: delivery is only half the work, while reviewed decisions and corrections become maintained input for later runs. It reinforces human judgment at planning and acceptance boundaries and narrows autonomy to workflows with dependable tests and review. Against broader skill systems, its key constraint is curation—memory must stay specific, current, retrievable, and evaluated rather than merely accumulate.

EveryInc/compound-engineering-pluginThe plugin turns Klaassen’s brainstorm-to-review loop and captured learning into a concrete, reusable six-stage repository workflow.SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient DescentSkillProx supplies measured retries, rollback, and pruning for the exact risk Klaassen identifies: accumulated lessons becoming noisy, stale, or harmful.Coding Agents Don't Scale Themselves. Neither Do Your Teams. — Patrick Debois, TesslDebois extends this solo compounding loop to organizations, where shared registries, evaluations, guardrails, and accountable platform owners are needed to maintain reusable context.Prototyping as Leadership: How a CTO Ships with AI Agents — Hursh Agrawal, The Browser CompanyThe leadership workflow reinforces Klaassen’s human-at-the-boundaries model through prepared context, bounded overnight execution, and personal testing and review before wider adoption.

Context Map

agentcoding#agent-memory#harness-engineering#skills

Uncertainty

The claim that implementation is mostly solved comes from one builder's workflow, not comparative evidence. Long autonomous runs still depend on strong evaluation, and accumulated memory can become noise unless lessons remain specific, current, and retrievable.