---
title: 'Human attention budget'
description: 'The finite supply of skilled human judgment available to specify, review, govern, and recover automated work.'
canonical_url: 'https://darkfactory.dev/glossary/human-attention-budget'
markdown_url: 'https://darkfactory.dev/glossary/human-attention-budget.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Human attention budget


## Definition

The finite supply of skilled human judgment available to specify, review, govern, and recover automated work. It includes focused time from people who can recognize a bad premise, resolve ambiguity, assess risk, approve exceptions, diagnose failures, and own the consequence.

AI changes the ratio between generated work and available review. Tokens and agent runs can scale quickly. Expertise, concentration, and organizational context do not. Sam Schillace describes this as an operating-model problem: attaching abundant AI generation to an unchanged workflow can overwhelm the scarce attention needed to decide whether the work is useful.

## Budget the decisions, not just the hours

A practical attention budget identifies which decisions truly need a person, what evidence lets routine cases proceed automatically, how uncertain cases are ranked, and when work must wait. The goal is not zero human involvement. It is to spend judgment where an error is costly or automation lacks the context to decide.

Attention also has quality and switching costs. Ten minutes from the accountable security owner cannot always be replaced by ten minutes from any available reviewer. Parallel agents may increase output while fragmenting the same person's concentration across too many live decisions.

## Distinguish it from nearby terms

A token budget limits model consumption. A review queue counts pending work. The human attention budget asks whether the right people can make the required decisions with enough context and focus. Human-in-the-loop is a control pattern; it is not proof that sufficient attention has been reserved.

## Check your understanding

Five agents each produce a pull request and request immediate architectural review from one staff engineer. Did parallelism increase throughput? Only if the review bottleneck and context switching do not erase the gain. Route routine evidence automatically, limit work in progress, and reserve expert attention for exceptions and irreversible decisions.

## Related terms

- [Token budget](https://darkfactory.dev/glossary/token-budget)
- [Token maxing](https://darkfactory.dev/glossary/token-maxing)
- [Cost per accepted durable outcome](https://darkfactory.dev/glossary/cost-per-accepted-durable-outcome)

## Related factory areas

- [Human roles, organization, expertise & cognitive health](https://darkfactory.dev/factory/human-roles-expertise)
- [Economics, capacity & factory FinOps](https://darkfactory.dev/factory/economics-finops)

## Evidence and further reading

- [Sam Schillace: AI is an Attention Mangler](https://sundaylettersfromsam.substack.com/p/ai-is-an-attention-mangler)
- [Agentic Coding and Persistent Returns to Expertise](https://www.anthropic.com/research/claude-code-expertise)
- [Collaborator or Assistant: Work Partitioning](https://arxiv.org/abs/2605.08017)
- [Why Software Factories Fail](https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/wsff.md)