Compounding Engineering and Human Attention

3. Enduring Principle

Convert repeated friction into evaluated reusable capability

Compounding Engineering is the practice of turning recurring, attributable human corrections and production outcomes into reviewed improvements to instructions, skills, tools, tests, context policies, workflows, documentation, or deterministic software.

The compounding loop operates below the broader continual-learning governance loop. It supplies concrete improvement candidates; it does not change active behavior by itself.

Harvest corrections with provenance and scope

A Correction Record should contain:

• source Attempt, artifact, and exact before/after behavior;
• human actor and role;
• reason, affected criterion, and confidence;
• correction type: factual, procedural, stylistic, architectural, policy, safety, or outcome;
• proposed scope: personal, repository, team, workflow, or organization;
• sensitivity, retention, and consent;
• recurrence evidence and related incidents; and
• candidate destination such as test, instruction, skill, tool, or code.

Do not learn from acceptance alone. A reviewer may accept under deadline, fix the result silently, or miss a defect. Capture explicit corrections and downstream outcomes.

Promote to the narrowest durable mechanism

Use the least probabilistic mechanism that solves the recurring problem:

An instruction is not automatically the best answer. Repeatedly telling an agent not to violate a structural invariant is weaker than making the invalid state impossible or testable.

Separate personal fit from organizational truth

A Human Workflow Profile may describe interaction preferences such as planning depth, increment size, review cadence, explanation style, preferred surface, notification channel, and accessibility needs. It must not grant tools, change policy, lower quality gates, or alter business authority.

Model and harness selection should consider human workflow fit in addition to task quality, cost, latency, security, and availability. Migration between models or harnesses should include training, paired use, opt-in canaries, workflow documentation, and measurement of correction and attention—not only a new default announced by benchmark score.

Name human control modes precisely

• Human-in-the-loop: a person performs a required decision or correction inside the workflow.
• Human-on-the-loop: the workflow operates within policy while a person supervises outcomes and handles exceptions.
• Human-out-of-the-loop: no human decision is required for that bounded workflow instance; prior human policy and accountability still apply.

These modes describe intervention frequency, not the removal of human accountability. A workflow can be out-of-the-loop for execution and still require human policy ownership, promotion, incident response, or risk review.

Budget human attention explicitly

Measure:

• time to first required human decision;
• decision and approval latency;
• correction and override rate;
• review minutes per accepted outcome;
• avoidable notification and false-escalation rate;
• exception age and ownership;
• time spent reconstructing missing context;
• repeated correction clusters; and
• cognitive load reported by users.

An Attention Budget defines the expected human effort for a workflow and which decisions justify interruption. An escalation should arrive as a decision packet with the affected outcome, risk, evidence, uncertainty, options, recommendation, deadline, and resume behavior.

Optimize small feedback increments when judgment is dense

For writing, design, architecture, or ambiguous product work, short iterative increments may create better calibration than one large generated artifact. The human supplies examples through edits; the agent applies the emerging pattern to the next bounded section. Once a pattern repeats, extract it into a reviewable style guide, anti-pattern catalog, example set, or skill.

For mechanical work with strong specifications and tests, larger autonomous increments may be appropriate. Interaction granularity is a workflow design choice, not a universal preference.

8. Notes and lessons learned

• The best reusable prompt improvement may be a test, tool, schema, or policy rather than more prompt text.
• Model-switching cost includes rebuilding human intuition and workflow habits.
• Attention is saved when the system presents a decision, not when it sends more activity notifications.
• Compounding should make the next correction less likely, not merely make the next generation longer.

9. Interview and discussion questions

1. What distinguishes compounding engineering from uncontrolled self-learning?
2. How should a personal preference become—or not become—a team rule?
3. When should a repeated correction become deterministic software?
4. How do human-in-, on-, and out-of-the-loop modes differ?
5. What would you measure to prove the factory reduces cognitive load?
6. How would you migrate a team between harnesses without losing productivity?