Compounding Engineering
Compounding engineering is the discipline of building each unit of work so that it makes the next unit of work easier. Coined at Every by Kieran and Nityesh; elaborated by Dan Shipper in Dan Shipper on Every.
Core principle
“For every unit of work we do, we try to make the next unit of work easier.”
Ordinary engineering produces output. Compounding engineering produces output and infrastructure that accelerates future output. The difference compounds over time.
Concrete mechanisms
| Mechanism | Description |
|---|---|
| Encoding feedback as prompts | After each review cycle, distil the feedback into a reusable prompt. Future work starts with accumulated judgment rather than from scratch. |
| Reusable automations | Build automations once; never rebuild them. Each automation reduces the marginal cost of the next similar task. |
| Head of AI operations | A role dedicated to converting tacit taste into explicit machine instructions. The role exists to make compounding systematic rather than accidental. |
| Knowledge bases | Invest in context assets (documents, memory stores) that the model can draw on — reduces per-query context reconstruction overhead. |
Relationship to context engineering
Compounding engineering and context engineering are adjacent but distinct:
- Context engineering (Tobi Lütke / Spotify): getting the right context to the model at the right time for any given query.
- Compounding engineering: building the system that accumulates and organises context so that context engineering becomes progressively easier.
Context engineering is a per-query problem; compounding engineering is the meta-problem of making those per-query operations cheaper over time.
See also Agentic Engineering for the correctness/oversight dimension, and AI Engineering for the system-design dimension.
Relationship to agentic engineering
Agentic Engineering focuses on correctness guarantees and oversight when AI agents generate code at scale. Compounding engineering is orthogonal: it is not about correctness but about throughput growth — designing the system so that each iteration leaves the system in a state that makes the next iteration faster.
The two disciplines are complementary: agentic engineering defines what counts as acceptable output; compounding engineering ensures that producing acceptable output becomes cheaper each time.
Head of AI operations
Dan Shipper identifies the head of AI operations as the role responsible for compounding engineering at the organisational level:
- Builds and maintains the prompt library.
- Translates the CEO’s feedback into reusable automations.
- Manages the knowledge bases that ground agent outputs.
Not a developer role. Requires communication skill, domain knowledge, and judgment about what the principal actually wants. See Dan Shipper on Every for Katie Parrott’s account of this role at Every.
The compounding dynamic
At Every (15 people, no manual code-writing, Cora built for ~$300K including salaries):
- Every feedback session becomes a prompt.
- Every prompt is accessible to the whole team.
- Young writers record all feedback sessions as structured prompts; Alex Duffy made a year’s progress in two months via this method.
The accumulation is not accidental — it requires deliberate engineering of the encoding process.