Organizations which design systems (in the broad sense used here) are constrained to produce designs which are copies of the communication structures of these organizations..
Background
In the first quarter of 2026, my social media feeds on LinkedIn, Twitter and XiaoHongShu are completely flooded by those around one single key word, agent. Much of this momentum can be attributed to the continuous productization and narrative-building efforts by OpenAI and Anthropic. This combination has made AI significantly more accessible, enabling people to directly leverage it to accomplish tasks, which is an undeniably important step forward.
However, as with any technological wave, rapid adoption often comes with a loss of perspective. Maintaining clarity becomes even more critical in enterprise digitalization, where the stakes are higher and the systems are more complex than individuals. Collaboration between human-and-agent as well as between agents is widely seen as the next frontier. Tools like minus, openclaw, hermes have already demonstrated how individuals can leverage agents to automate tasks at scale. In such scenario, it has been demonstrated that a single agent or a multi-agent system can complete tasks automatically and collaboratively - each agent is assigned a role, and its instructions are constructed based on that role’s responsibilities. There are lots of studies and researches in the field attempting to find the best technological paradigm to make the multi-agent system work effectively in an enterprise set-up. The interesting part is, once we move from single-agent capability to multi-agent collaboration in an organization with intensive human-to-human, human-to-agent, and agent-to-agent communications, a fundamental question emerges:
Collaboration is, at its core, a problem of communication.
In enterprises, communication is never free-form. Instead, it is always constrained by organizational structure. This is precisely what Conway’s Law captures: communication paths are shaped by how organizations are structured. Consider a typical enterprise with departments of procurement, supply chain, R&D, and operations, each forming distinct functional units. Procurement and supply chain, being closely related, tend to share systems and communication channels. In contrast, procurement and R&D, with weaker contextual overlap, often operate through entirely different systems. As a result, system architecture is fundamentally a projection of communication pathways, and organizational boundaries naturally become system boundaries. At a global level, inconsistencies are rarely caused by technical limitations, they are more often the result of communication costs. This becomes even more evident in the context of widespread agent adoption.
The agent setup in enterprise should favor the fundamental Conway’s Law, and this can be achieved in the following perspectives.
Communication Paths
The capability of modern AI agents is not only determined by model capacity, but also by the quantity and relevance of their context. When two teams share highly aligned context, communication becomes efficient, and systems built on top of that alignment tend to be reusable. In such environments, introducing multiple agents is significantly easier than those where context is diversified. This is because the communication pathways between agents, humans, and systems are already well-defined by the organization. The procurement team always check with supply chain team about inventory status whenever they place a new order; supply chain team will always rely on the procurement team for planning. Conversely, when teams lack shared context, deploying effective agent systems becomes difficult. The limitation is not the model itself, but the absence of a coherent communication structure.
Most of the agentic system assumes the structure of planner, executor, thinker, etc., which forms a natural workflow to accomplish tasks. However, this communication path does not tally with many of the organizational module configuration - maybe planning is not just the job scope of one single department but a shared task across many - the plan-then-execute workflow would then fail. The single route of plan-to-execute is against the communication path of the organization, and correspondingly, there will be much overhead to communication the context to proceed with execution as the next step.
The right approach is to architect the agentic systems with the role settings that maps the organizational communication paths. If the agentic system is meant for the collaboration between procurement and supply chain, then the agents should be designed to tackle the problems given the domain-specific instructions. That is, if the procurement agent just touched the inventory data, it can use the outcome to directly feed into the context of the supply chain agent. And this will not incur much overhead in communication.
Organizational Boundaries
Organizational boundaries manifest concretely through interfaces. For example, a procurement team may adopt a new system to manage inventory data. While this data is theoretically useful to the supply chain team, incompatibilities in data interfaces can prevent seamless collaboration, reducing overall efficiency. This issue persists in agent-driven systems. While agents can process unstructured data, they still rely on APIs, schemas, and data models, overall harness, to ensure reliable interaction. In other words, organizational boundaries do not disappear, they re-emerge as interface constraints.
Keeping clear interfaces making agentic systems or sub-systems work effortlessly. In such sense, agent look similar to microservices. For a microservice, the key components are the API, the localized memory, the resources to support the service runtime. And a similar setup can be found in the setup of agentic system - the only thing is that the APIs are enhanced as MCP and the database is enhanced as context or memory, to favour the interaction through human languages. Like what is seen in human organization, role-and-responsibility (R&R) is key to successful collaboration among agents. Defining the interface scope for these agentic systems is vital to a successful operation. Going back to the real-world example, the inventory data management can be largely relevant to both the procurement team and the supply chain team. Maybe there is shared database, context, memory, and MCPs for both of the two teams, but there should be clearly boundary definition as either wrapper of API, or routing layer of MCP, to make sure that when surfacing different teams, there are still clear boundaries in the R&R.
Communication Cost
All communication has a cost. For humans, this cost appears as meetings, fragmented documentation, and inconsistent understanding. For agents, communication is compressed into tokens, and the process is transformed into computation and uncertainty. Ideally, we want minimal tokens to yield precise and reliable outputs, enabling consistent propagation across systems. In reality, however, the probabilistic nature of large language models makes perfect accuracy unattainable.
Thus, reducing communication cost is not about eliminating uncertainty, but about constraining it. The prompt that is used for communicating among agents is structured as contract where the control framework is introduced. In such manner, the communication among agents become explicit, repeatable, and structured. That is, if the agent procurement team initiates a request to talk to the agent in the supply chain team regarding the inventory data related matters, the requests (or instantiated as prompts) will be through the checks against the shared context and long-term memory to guarantee that the requests are relevant with the minimal communication overhead. Meanwhile, the follow-up response or feedback from the supply chain team will help improve the context and memory shared across the two teams to make sure that the next round of communication is better constrained.
What can we do
Returning to the example above, where the enterprise has multiple functions, i.e., procurement, supply chain, operations, and R&D. Assuming each deploys agents to automate their workflows. A Conway-compliant architecture would resemble the diagram shown below. In this setup, each function operates through a combination of human workers and agents; each agent maintains its own local memory to preserve execution history, allowing it to remain coherent when receiving new context and instructions. Agents communicate with those in other functions through standardized interfaces, and during this process, shared context provides both the communication pathways and the boundary conditions that enable efficient collaboration.
As discussed earlier, shared context is the key enabler of smooth communication. However, in designing such systems, this shared context should not be expanded across the entire enterprise. Doing so would lead to “context explosion”, significantly increasing communication costs between agents—effectively a form of inverse Conway’s Law. Instead, shared context should be selectively established among functions with strong overlap and commonality, enabling efficient communication where it is actually needed.
Wrap-up
For organizations seeking to effectively adopt AI agents, understanding Conway’s Law is essential. Within a fixed organizational structure, success depends on:
- Defining clear context to shape communication paths.
- Establishing stable interfaces to enforce boundaries.
- Applying constraints to control communication cost.
Perhaps, we are entering a new era where systems mirror organizational structure is fading, and an era where systems reflect communication structure is beginning.
References
- Conway, M. (1968). How Do Committees Invent?
- Sussman, G. and Steele, G. (1975). Constraints and Communication in System Design.
- Brooks, F. (1975). The Mythical Man-Month.
Citation
Plain citation as
Zhang, Le. Conway’s Law in the Age of AI Agent Systems. Thinkloud. https://yueguoguo.github.io/2026/04/01/conway-law/, 2026
or Bibliography-like citation
@article{yueguoguo2026conwaylaw,
title = "Conway's Law in the Age of AI Agent Systems",
author = "Zhang, Le",
journal = "yueguoguo.github.io",
year = "2026",
month = "Apr",
url = "https://yueguoguo.github.io/2026/04/01/conway-law/"
}