Compound Engineering

Compound engineering is the discipline of turning engineering work into future engineering leverage. Each task should leave behind a durable improvement: a b…

Compound Engineering

Compound engineering is the discipline of turning engineering work into future engineering leverage. Each task should leave behind a durable improvement: a better plan template, rule, test, skill, reviewer, doc, workflow, or product insight.

Core rule

Every unit of engineering work should make subsequent units easier, not harder.

In ordinary codebases, features add complexity and future work slows down. In a compound system, bug fixes eliminate bug classes, plans become reusable, reviews generate new checks, and repeated manual work becomes tooling.

Durable artifact taxonomy

Every meaningful agent session should produce at least one of:

Artifact	Examples
Knowledge	wiki page, ADR, glossary entry, synthesis, product insight
Guardrail	regression test, lint rule, hook, eval, approval gate, security checklist
Capability	skill, subagent, script, workflow, reusable prompt, Graphify/qmd query pattern

If no durable artifact exists, the session helped once. If a durable artifact exists, the session compounds.

Loop

Compound engineering extends plan-first agent work:

1. Ideate: choose valuable work.
2. Brainstorm: define users, constraints, edge cases, and success.
3. Plan: research repo patterns and external docs.
4. Work: implement in isolation and validate continuously.
5. Review: use specialized reviewers and prioritize findings.
6. Polish: test the product experience, not just code correctness.
7. Compound: capture the reusable learning and update the system.

Safety tension

Compound engineering can either compound leverage or compound slop. Mario Zechner’s Pi talk is the counterweight: agents without learning, scope, and bottlenecks generate errors faster than humans can review them.

The synthesis:

• Parallelize scoped, measurable, non-critical, or well-verified work.
• Keep humans in the loop for critical judgment and product direction.
• Replace manual review bottlenecks with safety nets only when the safety net is trustworthy.
• Cap generated code to what can be reviewed or objectively verified.

Fit with this wiki

llm-wiki is already a compound engineering system:

• raw/ and pdfs/ are source inboxes.
• wiki/ compiles sources into durable concepts and syntheses.
• qmd retrieves prior learnings just in time.
• mistakes/ converts agent failures into future rules.
• AGENTS.md / CLAUDE.md codify current taste and process.
• Graphify is a local architecture graph for codebase questions when graph context helps.

Related

• Compound Engineering — source summary.
• Compounding Knowledge Base — same principle applied to knowledge compilation.
• Agent Harness — runtime substrate for long-running agent work.
• Verification Pipeline — safety-net side of compounding.
• Agent Context Instructions — taste encoded as durable instructions.
• Pi Building In World Of Slop — warning against unbounded agent output.