Leverage Laws: A Per-Task Framework for Human-Agent Collaboration

1. Problem Statement

Current literature on human-AI collaboration has established that AI improves output, that productivity gains vary by task, and that autonomous time horizons are increasing. However, the field lacks a unified unit of analysis to answer the operational question: Where should a practitioner invest the next hour of work to maximize the rate at which agents displace human labor?

Existing frameworks are qualitative (e.g., supervisory control levels, common ground theory) or focus solely on aggregate output. There is no normative ratio that decomposes the cost of collaboration into measurable channels, allowing for the optimization of workflow design versus model capability.

2. Methodology

The paper proposes a normative framework centered on a per-task leverage ratio ($L_{task}$) and extends it to a windowed leverage measure ($L_{window}$).

A. The Base Formula: Per-Task Leverage

The core metric is defined as the ratio of human work displaced to the human time invested:
$$L_{task} = \frac{H_{displaced}}{t_{planning} + \sum t_{interrupt_j} + t_{review}}$$

• Numerator ($H_{displaced}$): Human-hours removed from the operator's plate by the agent (governed by agent capability).
• Denominator: Human-hours spent on three distinct channels:
  1. Planning ($t_{planning}$): Specifying the task.
  2. Interrupts ($\sum t_{interrupt_j}$): Resolving mid-run issues.
  3. Review ($t_{review}$): Verifying output.

B. Information Density and Directional Asymmetry

The framework introduces Information Density ($\rho$) as bits per hour, decomposed into two directional flows:

• $\rho_{in}$ (Human-to-Agent): Bottlenecked by human output (speech/typing). Asymptote is $\approx 20$ bits/sec.
• $\rho_{out}$ (Agent-to-Human): Bottlenecked by human input (reading/visual parsing). Has significantly higher headroom due to the agent's ability to choose high-density modalities (e.g., dashboards vs. prose).

The effective density of an exchange is the harmonic mean of these two rates, weighted by the fraction of information flowing in each direction ($\alpha$). The paper argues that different phases of collaboration have different $\alpha$ values:

• Planning: High $\alpha$ (input-heavy).
• Interrupts: Balanced $\alpha$ (bidirectional).
• Review: Low $\alpha$ (output-heavy).

C. Information Conservation Law

Every task has a fixed information requirement ($I_{task}$) that must be conveyed. This information flows through the three channels. The framework posits that under-specifying in planning forces information into interrupts or review, where the "cost scalars" ($c_p, c_i, c_r$) are often higher due to context-switching and rework.

D. Asymptotic Analysis

The paper analyzes the behavior of $L_{task}$ as system capabilities and shared memory ($M$) approach infinity:

• Interrupts and Review: These terms collapse to zero as memory improves and verification becomes automated.
• Planning: Approaches a non-zero floor determined by irreducible task novelty ($I_{novel}$). Even with perfect memory, new tasks require unique intent specification.
• Result: $L_{task}$ is bounded by the ratio of displaced work to task novelty, scaled by human input bandwidth ($\rho_{in}$).

E. Windowed Leverage ($L_{window}$)

To address recurring tasks and system design, the authors introduce a time-windowed measure. This allows for amortization:

• Recurring Tasks: Planning costs are paid once and divided by $N$ runs.
• Spawned Subtasks: Children inherit parent memory, reducing their planning costs.
• System Design: Upstream investments (templates, skills) amortize across all future tasks.
  Unlike $L_{task}$, $L_{window}$ has no per-task ceiling but is bounded by the stock of accumulated planning investment.

3. Key Contributions

1. Unified Normative Ratio: A single formula ($L_{task}$) that operationalizes three previously disjoint concepts: supervisory control costs, common ground (shared context), and mixed-initiative interruption logic.
2. Directional Information Asymmetry: A theoretical proof that human-to-agent and agent-to-human flows have distinct bottlenecks and ceilings. This implies that improving input modalities (e.g., voice) primarily helps planning, while improving output modalities (e.g., visualization) primarily helps review.
3. The "Irreducible Novelty" Floor: The identification that planning time cannot be eliminated, only compressed, because every task contains a unique information component ($I_{novel}$) that cannot be predicted by memory.
4. Amortization Framework: A shift from per-task analysis to per-window analysis, explaining how recurring tasks and system design investments break the per-task leverage ceiling.
5. Falsification Protocol: A specific, "Popper-grade" experimental design to test the directional asymmetry claim by measuring time deltas when $\rho_{in}$ and $\rho_{out}$ are independently improved.

4. Results and Theoretical Findings

• Decoupling of Levers: Investments in model capability (numerator) and workflow design (denominator) are first-order decoupled but co-evolve at the second order.
• Scaling Axes: Leverage scales along two axes: Capability (increasing $H_{displaced}$) and Memory (increasing $\rho$ and reducing $t_{interrupt}$ and $t_{review}$).
• Boundedness:
  • $L_{task}$ is bounded by $H_{displaced} \cdot \rho_{in}^{max} / I_{novel}$.
  • $L_{window}$ is bounded only by the accumulated stock of planning investment available within the time window.
• Phase-Specific Optimization: The framework predicts that optimizing a specific phase (e.g., review) requires targeting the specific information flow direction dominant in that phase (e.g., $\rho_{out}$ for review).

5. Significance

• Operational Guidance: Provides practitioners with a concrete method to diagnose bottlenecks. If $t_{review}$ is high, the solution is better agent output modalities ($\rho_{out}$), not better human input.
• Strategic Investment: Distinguishes between investments that raise the ceiling of leverage (improving $\rho_{in}$ for novel tasks) and those that increase the rate of leverage (amortizing costs via system design).
• Research Roadmap: Identifies 10 open empirical questions, including the functional form of memory density $\rho(M)$, the measurement of cost scalars ($c$), and the dynamics of recursive task hierarchies.
• Theoretical Synthesis: Bridges the gap between qualitative HCI theories (supervisory control, common ground) and quantitative productivity metrics, offering a rigorous mathematical basis for the future of human-agent collaboration.

In summary, the paper moves the field from asking "Does AI help?" to "How do we mathematically optimize the flow of information between human and agent to maximize labor displacement?" by defining the fundamental limits and levers of that interaction.