Compound Engineering & the Dark Factory

Two opposing answers to the same question: what does engineering become when agents write the code? **Compound engineering** keeps the human at the high-leverage ends of the loop and makes the system measurably smarter every cycle. The **dark factory** removes the human from the deployment path entirely and lets specs flow to production in the dark. This module is the synthesis of the whole course — it composes the harness (EB-2), evaluation (EB-6), guardrails and audit (EB-8), and memory (EB-5) into a working methodology and a working pipeline, and it closes the human-in-the-loop thread that runs through EB-1, EB-7, EB-8, and EB-9.

Theme: Foundations →

Topics

13 topics — scope and references for each.

The compound principle — work that makes future work easier
Compound engineering inverts the usual entropy of software. In traditional development each feature tends to make the next one *harder* — more code, more edg…
The compound loop — plan, work, review, compound
The methodology runs as a repeating loop that splits work into three phases: a human decides *what* is worth building, the agent plans/codes/tests/reviews an…
Levels of agentic autonomy
A ladder for locating where a developer or team currently operates, borrowed in spirit from the self-driving-car levels. It runs from manual coding, through…
What a dark factory is — and how it differs from a harness
The term comes from "lights-out" manufacturing — a plant that runs with no humans on the floor, the canonical example being a robotics maker whose robots bui…
Codifying learnings — instruction files, solution docs, the compound step
The mechanism that makes compounding real rather than aspirational. After each cycle the team captures what worked as durable, retrievable knowledge: short a…
Spec-driven input — plans and specs as the primary artifact
Both methodologies relocate the engineer's effort upstream of the code. In compound engineering "plans are the new code": the plan document becomes the sourc…
Building the compound system — parallel review, the 50/50 rule, agent-native parity
The production practices that make the loop scale. Students learn three load-bearing ideas. **Parallelisation:** the bottleneck is no longer human attention…
Dark factory architecture — the four layers and generator/evaluator isolation
The reference architecture has four layers: an **inputs** layer owned by humans (specs, holdout scenarios, instruction files, linter rules); an autonomous **…
Holdout scenarios and the train/test wall
The trust mechanism at the heart of the dark factory, borrowed directly from how ML models are evaluated. Acceptance criteria are written as plain-language b…
Phased rollout and trust gates
Autonomy is earned, not declared. Students learn a phased path that pays off at every stop: first give agents better context (instruction files, build-before…
Dark-factory failure modes and maintenance agents
Where autonomous pipelines break. The headline risk is **evaluator overconfidence** — an evaluator that keeps approving subtly flawed code because the criter…
Economics, leverage, and organisational redesign
What changes for the business, not just the codebase. Students examine the reported economics — small teams claiming the sustained output of teams several ti…
Identity, autonomy, and the human as metacognitive controller
The human and cognitive dimension of handing the work to agents, and the module's capstone. Students engage with the documented psychological barriers — lett…

Topics are combined ad hoc for corporate programs and workshops — there is no fixed public duration. Tell us your goals and we design the track.

Discuss a custom program