Gå til indhold
Use casesLærOm mig
cleverest
Library

Compound engineering

Compound engineering

Compound engineering is the practice of building AI workflows and automations that accumulate value through reuse and refinement. Rather than solving each task from scratch, compound engineering creates reusable systems — agents, automations, and integrations — that improve each time they're applied. The name reflects the compounding nature of the investment: initial effort building a workflow pays dividends across all future uses.

The approach represents a shift from ad-hoc AI assistance to systematic capability building. Instead of prompting for help with individual tasks, compound engineering creates persistent infrastructure that handles categories of work.

The compounding mechanism

When a workflow is built once and reused across projects, two forms of value compound. Time savings accumulate with each use — the upfront investment in designing the workflow amortizes across all executions. And the workflow itself improves through iteration: each use reveals edge cases, suggests refinements, and builds institutional knowledge about how to handle specific situations.

This creates asymmetric returns on effort. A few hours invested in building a robust workflow might save hundreds of hours across future projects while producing consistently higher quality than ad-hoc approaches.

Impact on work quality

A counterintuitive finding from compound engineering practice: automation doesn't just speed things up — it actually improves output quality. When repetitive tasks no longer consume bandwidth, attention shifts to work that was previously squeezed by time constraints: strategy, research, learner understanding, realistic scenario development. These high-value activities directly improve final output quality in ways that rushing through busywork never could.

The equation isn't "same work, done faster." It's "better work, because more attention goes where it matters."

Components of a compound engineering system

Effective compound engineering typically involves multiple components working together. Agents handle specific task types — collecting material, structuring content, reformatting drafts, managing reviews. Automations connect triggers to actions, handling routine flows without manual intervention. Integrations bridge different tools and data sources, creating unified workflows across previously siloed systems.

These components rarely stand alone. A well-designed compound engineering system chains multiple elements: an integration pulls source material, an agent processes it according to established patterns, automation routes the output to appropriate destinations. Each piece is individually simple; the compound value emerges from orchestration.

Building for reuse

Creating genuinely reusable workflows requires thinking beyond immediate needs. Documentation matters — a workflow that only its creator understands can't compound across team members or future contexts. Generalization matters — workflows too tightly coupled to specific projects don't transfer well. Maintenance matters — workflows require upkeep as underlying tools and requirements evolve.

The investment profile differs from one-off solutions: more effort goes into design and documentation upfront; less effort goes into each subsequent execution. This tradeoff only pays off when the workflow genuinely sees repeated use, making careful selection of what to systematize an essential skill.

Continuous improvement loops

The most powerful aspect of compound engineering is systematic improvement over time. Every workflow benefits from a feedback mechanism that captures insights from each execution and folds them back into the system.

The before-and-after technique exemplifies this. When output from an AI system requires manual correction, that correction represents valuable signal. Rather than discarding this information after fixing the immediate problem, a continuous improvement loop captures both the original output and the corrected version, then extracts rules or guidelines that prevent similar issues in future runs.

The key insight is minimizing friction in the feedback path. If improvement requires significant manual effort, it won't happen consistently. The most effective loops embed feedback collection into natural workflow touchpoints: comments added while reviewing output, corrections made during normal use. A fitness protocol integrated with Todoist, for example, allows feedback through task comments the user would write anyway. An email drafter captures corrections simply by comparing drafted and sent versions. When improvement happens as a byproduct of normal work rather than a separate activity, compound engineering reaches its full potential.

Visibility compounds capability across an organisation

A second compounding mechanism appears at the organisational level. When AI workflows operate in shared channels rather than private interactions, every successful execution serves as a small demonstration to everyone watching. Colleagues learn what kinds of work can be delegated to which workflow, build trust through observation rather than through individual evaluation, and absorb working patterns by scrolling back through channel history.

A perfectly designed automation buried in someone's private terminal helps that one person; the same automation visible in a team channel teaches the team. Defaulting to shared spaces — team channels, public documents, group conversations — produces faster organisational adoption than private interactions, even when per-task quality is identical. The compounding effect depends on the visibility, not on the workflow being any better than its private counterpart.

Related pages