📈 economics

The Organizational Behavior of Agentic AI: Collective Intelligence in Human-Agent Workflows

This paper argues that agentic AI functions as a partial organizational analogue to human teams, where collective intelligence is sustained not by social factors like trust and identity but by "context architecture," ultimately demonstrating that shared-state and adaptive forms outperform human-imitation structures by minimizing transaction costs through durable, inspectable, and task-contingent contexts.

Original authors: Canhui Liu

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Canhui Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are running a busy kitchen. For a long time, you've had human chefs. They have personalities, they get tired, they argue, they trust each other, and they know the "unwritten rules" of the restaurant.

Now, imagine you replace those chefs with a team of AI robots. But here's the twist: you don't just have one robot chopping vegetables. You have a whole crew: a "Planner" robot that writes the recipe, a "Solver" robot that cooks, a "Reviewer" robot that tastes the food, and a "Memory" robot that remembers what ingredients you have.

This paper asks a big question: Is this robot kitchen actually a "team" in the same way a human kitchen is?

The author, Canhui Liu, says the answer is "Yes, but..." It's a "partial copy."

Here is the breakdown of the paper's main ideas using simple analogies:

1. The "Yes" Part: They Look Like Teams

Just like a human team, these AI robots have to do the same things to get work done:

Splitting the work: The Planner breaks a big task (like "make a lasagna") into small steps.
Passing the baton: The Solver passes the half-cooked dish to the Reviewer.
Following a routine: They do the same steps over and over again.
Crossing boundaries: They move information from the "memory bank" to the "cooking station."

Because they do these things, we can use old business theories to understand them. They act like a team.

2. The "But" Part: They Are Not Human

This is where the paper gets interesting. While the robots act like a team, they don't feel or think like one.

No Trust or Identity: A human reviewer might say, "I trust my colleague's cooking," or "I'm a professional chef, so I won't serve bad food." The AI reviewer doesn't have feelings, pride, or a career. It only does its job because its instructions (prompts) tell it to.
No "Social Cost": If a human chef makes a mistake, you might fire them or they might feel bad. If an AI robot makes a mistake, you just reset it and try again. It doesn't cost you any "social" money.
The "Black Box" Problem: Humans have a shared understanding of the world. AI robots only have context architecture. This is a fancy way of saying they rely on the specific notes, memory logs, and rules you give them. If you forget to write a note in their "memory," they forget the whole task.

3. The Big Problem: "Context Transaction Costs"

The paper introduces a new idea called Contextual Transaction Cost.

Think of this like passing a note in class.

In a human team: If I tell you a secret, you understand the tone, the urgency, and the context. You get it immediately.
In an AI team: To pass the "secret" from the Planner to the Solver, the system has to write it down, compress it (summarize it), send it, and then the Solver has to read it and guess what it means.

Every time the robots pass a task to each other, they lose a little bit of information (like a game of "Telephone").

The Cost: If the robots have to pass the ball back and forth too many times, or if they summarize the instructions too much, they lose the "context." They start making mistakes because they forgot the original goal.
The Paper's Finding: Sometimes, having a "Manager" robot to coordinate things actually makes things slower and worse because it adds too many handoffs. Sometimes, it's better if the robots just share a single "whiteboard" (shared memory) so they don't have to pass notes back and forth.

4. The Solution: The "Interface"

The paper argues that we shouldn't try to make AI act exactly like humans (that's a trap). Instead, we need to build a special bridge between the human world and the robot world.

The Human Side: Humans need to know why the robot made a decision. They need to see the "receipt" (the trace of what the robot did).
The Robot Side: The robots need clear rules on what they are allowed to do and what information they must keep.

The paper suggests that the best way to use these AI teams isn't to copy human management styles (like having a "Committee" of robots). Instead, we should design systems where the robots keep their "notes" (evidence and traces) very clear so humans can check their work easily.

Summary

The paper says: AI teams are like a very efficient, very obedient, but very fragile assembly line.

They are great at breaking down big jobs and doing them fast.
They are terrible at "guessing" what you meant if you didn't write it down perfectly.
If you treat them like human employees (expecting them to "trust" each other or "take initiative" without clear rules), they will fail.
If you treat them like a machine system where the flow of information (context) is carefully managed, they can be incredibly powerful.

The goal isn't to make robots human; it's to build a workflow where human judgment and robot efficiency fit together without losing the "notes" in the middle.

Technical Summary: The Organizational Behavior of Agentic AI

Problem Statement
Agentic artificial intelligence is increasingly deployed not as isolated tools but as collectives of planners, solvers, reviewers, and orchestrators that mimic human organizational structures (e.g., teams, committees, markets). While these systems are described using organizational terminology, a critical gap exists in understanding whether they truly exhibit "organizational behavior" comparable to human collectives. Current discourse risks two errors: instrumentalism (treating agents as neutral tools) and anthropomorphism (assuming agents function like human teams with identity, trust, and moral accountability). The central problem is determining how agentic AI behaves organizationally, how it differs from human organizations, and under what conditions human-agent workflows can achieve collective intelligence without incurring systemic dysfunction.

Methodology
The paper employs computational theorizing, a strategy that formalizes mechanisms and varies operating conditions to observe resulting patterns. The empirical evidence is drawn from three distinct sources:

Synthetic Simulation (Study 1): A large-scale simulation of 8,000 knowledge-work tasks evaluated across seven distinct agent organization forms. This allows for controlled variation of task decomposability, error correlation, and communication modes.
Trace-Instrumented Real LLM Runs (Study 2): Empirical analysis of actual Large Language Model (LLM) executions across four domains: software repair, long-document question answering, legal review, and literature synthesis. This study focuses on estimating "contextual transaction costs" within real agent traces.
Robustness Analysis (Study 3): Validation using open-source models, prompt variations, and a prompt-free reinforcement learning (DQN) organization selector to ensure findings are not artifacts of specific prompting or model architectures.

The research design treats agent traces as observable "patterned action," allowing for the measurement of organizational dynamics (differentiation, coordination, boundary crossing) that are often opaque in human organizations.

Key Contributions and Theoretical Framework
The paper proposes that agentic AI is a partial organizational analogue. It shares functional similarities with human organizations (differentiation of work, coordination of interdependence, recurrent routines, and boundary crossing) but lacks the social foundations (motivation, identity, trust, employment relations) that sustain human behavior. Instead, agentic behavior is sustained by context architecture (prompts, memory, traces, schemas, tools, validators, and permissions).

The central theoretical mechanism introduced is Contextual Transaction Cost (CTC). Extending Coasean transaction cost theory, CTC defines the cost of moving task context across boundaries within an agent collective. It is modeled as a function of:

Token and latency burdens ( $\tau$ )
Cross-agent handoffs ( $H$ )
Compression loss ( $C$ )
Semantic drift ( $D$ )
Verification burdens ( $V$ )
Governance costs ( $G$ )

The paper argues that collective efficiency is achieved only when marginal gains from specialization, parallelism, diversity, and verification exceed the marginal costs of these contextual transactions.

Results and Findings
The empirical analysis yields several critical insights regarding the performance of different agent organization forms:

Failure of Human Imitation: Organization forms that strictly mimic human structures (e.g., free-text debate committees, sequential handoffs) often underperform. They introduce "lossy handoffs" where context is compressed or lost, leading to correlated deliberation and verification burdens that outweigh the benefits of division of labor.
Success of Shared-State and Adaptive Forms: Organization forms that prioritize shared state, durable context, and task-contingent adaptation perform better. These forms minimize semantic drift and ensure that evidence remains attached to claims across the workflow.
The Illusion of Pseudo-Diversity: Increasing the number of agents does not guarantee improved performance if errors are correlated. "Free-text debate" scales poorly unless agents utilize independent models, tools, or evidence sources.
Context Durability: High-quality outputs are less valuable if the underlying context is opaque or fragile. In legal and engineering workflows, an output with slightly lower quality but strong evidence preservation and inspectable traces is often a superior organizational outcome.

Significance and Claims
The paper claims to contribute to organization studies by theorizing agentic AI as a distinct object of analysis: an emerging object of organizing. It does not claim that agentic AI replaces human organizations or that it possesses human-like agency. Instead, it specifies the interface conditions required for human-agent collaboration.

The primary significance lies in the concept of the Interface Organization: the designed arrangement that aligns human accountability with agentic context coordination. The paper argues that for collective intelligence to emerge in human-agent workflows, the system must ensure that agent outputs are accompanied by preserved traces, disclosed uncertainty, and appropriate permissions. The work shifts the focus from "model quality" to "organizational design," suggesting that the efficiency of agentic AI depends less on the raw capability of individual models and more on how context is managed, preserved, and transferred across the collective.