Composition

In plain language: The word is plain — no separate name needed.

Definition

Composition is what happens when Agents combine. Two or more Agents working together, interacting, or coupled in any arrangement form a Composition. The Composition has its own three Boundaries — Computational, Cognitive, and Causal — that are not found by adding up the Boundaries of the individual Agents.

This is the framework’s central claim about combining: Composition is not additive. No single person can build a 747. Assemble thousands of people into a company with the right structure, and the collective can — the Composition’s Cognitive Boundary contains “how to build a 747” even though no individual member’s does. Conversely, a committee of smart people can produce decisions dumber than any individual member would have made alone. The same act of combining can expand one Boundary while contracting another.

The framework names four outcomes that describe what combining does to each Boundary:

• Expansion — the Composition can sense, model, or affect things that no individual member can. A honeybee swarm selects a nest site more accurately than any individual scout.
• Redundancy — multiple Agents with overlapping Boundaries on a given dimension. Sometimes waste; sometimes robustness. Two eyes do not just see more — their overlap creates depth perception, a qualitatively new capability from redundant input.
• Contraction — the Composition is less capable than its best individual member on that Boundary. Groupthink, information cascades, diffusion of responsibility.
• Reshaping — the Composition’s Boundary is differently shaped, not clearly larger or smaller. A committee of specialists may gain specialist depth while losing generalist breadth.

These are descriptions, not types. They apply per Boundary, not per Composition. A single Composition can show Expansion on the Cognitive Boundary, Contraction on the Causal Boundary, and Reshaping on the Computational Boundary — all at the same time. The words describe the shape of the change on each Boundary independently.

What determines which outcome appears on which Boundary is the framework’s primary research question — the composition problem. The answer sits in how the Agents are arranged (Topology), what rules govern their interaction (Protocol), and how well their Boundaries couple at the interfaces where one Agent’s Causal output becomes another Agent’s Computational input. Each of these is defined downstream. The composition problem is the reason the framework exists; no single definition can answer it.

What Composition is not:

• Not the sum of individual Boundaries. If it were, there would be no composition problem — you would just add up what each Agent can sense, model, and do. The non-additivity is what makes the question interesting and the framework necessary.
• Not determined by a single factor. Topology, Protocol, member capabilities, redundancy, diversity, delegation dynamics — all contribute, and no single one dominates universally.
• Not inherently good or bad. Expansion is not always desirable (expanding the Causal Boundary of a poorly governed Composition increases the blast radius). Contraction is not always a failure (narrowing individual discretion through Protocol can prevent harmful Causal output). The Capacity is not Value principle applies to Compositions as it does to individual Agents.

Relations

Composition is where the framework’s three Boundaries meet the multi-Agent case. Every Boundary defined in the first twelve terms applies to individual Agents. Composition is where those individual Boundaries combine into collective Boundaries with their own geometry.

A Composition that meets the Agent test — it senses, models, and acts — is itself an Agent at a larger scale. Per Turtles All The Way Down, the same three-Boundary analysis applies at the Composition level as at the individual level. A company is a Composition that is also an Agent. A team within that company is a smaller Composition that is also an Agent. The framework applies recursively.

Organization is the common term for the multi-Agent Composition case — a company, a team, an immune system, a multi-agent deployment. Topology describes how the Agents within a Composition are arranged. Protocol describes the rules governing their interaction. Composition Profile is the diagnostic tool for analyzing a specific Composition — a matrix that maps the four outcomes across all three Boundaries at both the Capacity and Realized layers.

The gap between a Composition’s Capacity and its Realized output is one form of Trapped Intelligence. The three-Boundary framework supplies per-failure-mode diagnosis: if the Composition has the information but cannot model it, the failure is Cognitive; if the Composition can model it but cannot act on it, the failure is Causal; if the Composition never received the information in the first place, the failure is Computational. Different failures, different interventions.

Frame applies at the Composition level. A Composition’s collective self-model — what the group thinks it can sense, model, and affect — can diverge from the actual collective Boundaries. A management team that believes it understands its market when its Cognitive Boundary does not extend to the competitive dynamics reshaping it is the multi-Agent version of the same Frame-Boundary gap that operates at the individual level.

Example — CEO

A mid-size software company has a sales team and a product team. Each team is itself a Composition. When the two teams work together, the company-level Composition has Boundaries that neither team has alone.

On the Causal Boundary: Expansion. The sales team can sell but cannot decide what gets built. The product team can shape the roadmap but has no direct customer relationships. Together, the company can identify what customers need and build it — a Causal reach that neither team possesses individually.

On the Cognitive Boundary: Contraction. The sales team talks to customers every day. Individual reps know which features customers want, which competitors are gaining ground, and which accounts are at risk. This knowledge exists inside the Composition — it has crossed the sales team’s Computational Boundary and been modeled at their Cognitive Boundary. But the product team never sees most of it. The weekly cross-functional meeting covers three bullet points per customer; the nuance disappears. The company’s collective Cognitive Boundary on customer needs is narrower than the sales team’s alone — the Composition contracts Cognitive reach because the coupling between the teams does not carry the signal with enough fidelity.

The CEO who sees only the Causal Expansion (“we ship product”) misses the Cognitive Contraction (“we ship the wrong product because the product team does not know what customers actually want”). The Composition Profile is designed to surface exactly this: what happened to each Boundary when the Agents combined? The CEO can expand the Causal Boundary through composition while simultaneously, invisibly, contracting the Cognitive Boundary through the same composition. Seeing both at once is the diagnostic value.

The cause of the Cognitive Contraction sits in the coupling between the two teams. The sales team’s Causal output (what they communicate) does not carry the full content of their Cognitive Boundary (what they know). The product team’s Computational input (what reaches them through the weekly meeting) is a lossy compression of what the sales team could share. The intervention is at that interface: a better meeting format, a shared artifact, a direct channel between the sales rep and the product manager working on the relevant feature. This Contraction is not automatically a problem. The compression may be deliberate — not every customer request should reach the product roadmap, and filtering is itself a valuable function. The net effect on the company as a whole may be better with this Contraction than without it. Whether the gap between what sales knows and what product sees is worth closing depends on what the company is trying to accomplish — the Capacity is not Value principle applies to Compositions as it does to individual Agents.

Example — Research

A multi-agent AI system deploys two agents: a planning agent that receives a user’s request, breaks it into sub-tasks, and decides the order of execution; and an execution agent that takes each sub-task and carries it out — writing code, calling APIs, producing files. Each agent is an Agent in the framework’s sense. Together they form a Composition.

On the Causal Boundary: Expansion. The planning agent can reason about complex multi-step tasks but cannot write code or call APIs. The execution agent can write code and call APIs but cannot decompose a complex request into a coherent sequence. Neither can fulfill the user’s request alone. The Composition can — the collective Causal Boundary includes both planning and execution, a reach that exceeds either individual.

On the Cognitive Boundary: Reshaping. The planning agent has broad domain knowledge and can model dependencies between sub-tasks. The execution agent has deep knowledge of specific APIs and code patterns. Together, the Composition’s Cognitive Boundary is differently shaped — broader on planning, deeper on execution — without a clear net gain or loss.

The coupling failure is where the Composition reveals its vulnerability. The planning agent produces a sub-task description. The execution agent interprets it. If the description is ambiguous — the planning agent’s Causal output does not fully convey its Cognitive model of what the sub-task requires — the execution agent acts on an incomplete understanding. The Composition’s collective Causal output diverges from what the planning agent intended. The planning agent modeled the right thing; the execution agent did the wrong thing; neither is individually at fault. The failure lives in the interface between them — the Causal-to-Computational coupling that connects one agent’s output to the other agent’s input.

This is Trapped Intelligence at the Composition level. The planning agent has the knowledge. The execution agent has the capability. The Composition has both — at Capacity. But at the Realized layer, the knowledge does not reach the capability with enough fidelity. The same structural phenomenon the CEO faces with his sales and product teams: the information exists inside the Composition, the capability exists inside the Composition, and the coupling between them is where the value gets lost.