A Guide to AI Agents: Capturing Relationships, Processes, and Roles

Developing multi-agent systems leads naturally to designing org charts first, but org charts are static, hierarchical depictions of work, while work itself is much more dynamic. In the long run, an employee may work as a project manager in one workflow, a reviewer in the next, and play a supportive role in yet another workflow. The same can be said for software tools that can serve multiple purposes depending on the workflow context. The organization captures how work gets done by depicting workflows rather than reporting lines. Using this perspective provides a much closer representation of how the work that leads to productivity, in terms of tasks, responsibilities, and collaborations is actually being captured.

Capturing this activity requires some way documenting the processes that define what work looks like from day to day. Examples of daily processes can include onboarding a new employee, processing a reimbursement, or resolving a ticket. These processes help in defining and recording not just the process but the roles that the human, software, or agent will be implementing. More directly, the organization needs to define the links that convey who completed which tasks, and if any, are handing work off to someone else in the workflow. This fluid map of organizational activity will create the foundation for multi-agent systems as the agents are now encapsulating workflows, and doing so in meaningful ways. 

Today, technology offers the ability to map workflows like never before. And we have previously discussed the virtues of process mining software and conducting interviews to record undertones and nuances. However, generative AI offers the prospect of effectively discovering workflows with tremendous speed. By providing the AI with one sufficient description of a company or industry, the generative AI can parse out the language and define roles, processes, and connections and output in hours a structured and augmentable definition of an agent network. This ultimately means it may take days instead of months to prototype agent ecosystems versus traditional analysis. Again, however, in all of the speed and capabilities presented by new tooling options, one design principle should never be overlooked as we design the agent network as a Directed Acyclic Graph, otherwise organizations run the risk of creating infinitely looping agents delegating tasks back to each other, and back again. In one case, if an HR agent delegated a query to a Finance agent, it is entirely possible for that Finance agent to pass the original query back to HR and now our HR agent is stuck processing requests that are linking back to itself. Enforcing a DAG structure keeps things stable and predictable; it is the same reason we don’t have a system called the “Infinite Loop.” 

Agent Autonomy Calibration 

Within the organizational setting, not all tasks should be automated to the same degree. The development of effective agents requires calibration of autonomy. In some domains—such as law, medicine, and finance—the risks of making errors are too great to permit agents to act autonomously. In these domains, agents can best support human professionals as advisors, with human professions responsible for the final decision-making. For example, an agent responsible for drafting contracts may propose clauses with final approval by a person legally trained to draft contracts.   

Some settings look more like partial autonomy for agents. Agents may function autonomously with the understanding that they will escalate exceptions to human professionals. For example, an agent responsible for expense approvals may process approval for small items and then escalate medium refuge approvals to a manager for review. Fully autonomous agents are best, when in situations where speed:

Outweighs caution and an agent’s judgement is acceptable or required. An example of an agent performing full autonomy is an agent that monitors cloud services that spin up new servers automatically during a spike in cloud traffic. Finding the appropriate balance between zero autonomy, partial autonomy, and full autonomy is the best way to create a trustworthy multi-agent system and sustain adoption.   

Co-building Specialized and Third-Party Agents 

The AI application market is already crowded with specialized agents. For example, a coding agent, GitHub Copilot, provides support when coding specific programming languages. The example of legal AI refers to agents specifically designed to assist in reviewing legal contracts. Customer support bots are other examples of specialized agents. There are advantages to integrating third-party agents from an AI marketplace (e.g., Google Cloud AI), rather than recreating capabilities within the agent network you are building. Agents provide a wrapper to interact with third-party agents through APIs, admin support (i.e., file a form), and request a user-standardize input output format to analyze the modular node of specialized agents in the multi-agent agent application framework. This plug-and-play functionality will allow for improvement, reduce your time-to-market, and allow any integration of capabilities to be “future-proofed,” since you are not replacing the third-party agent, but simply integrating any new agent agency that a third party will provide.

Tools and Roles Inside Agents

Agents are not simply people solving problems—they need actual tools and capabilities. For example, a product manager agent might have a Kanban board act upon to promote a task being tracked, or an alerts agent sends real-time alerts to stakeholders. An intent agent will likely query a structured database—structuring the data and presentation itself would be a key capability for a database agent. Frequently, down-chain agents will also become the tool themselves allowing for modular and malleable opportunities for agents to work together. This propels the mindset of an agent ecosystem to a toolbox of generalized specialties and remarkable capabilities, and is now only dynamically orchestrated.

The Current State of Multi-Agent Systems

Current multi-agent systems, such as Microsoft AutoGen, serve as illustrations of the potential of what can be done. However, existing solutions are typically owned by the market (off the shelf) for reasons of scalability but present challenges because they are deployed as rigid, hard coded agents. As such, the more scalable approach is to build a model whereby down-chain agents are treated as tools, to extend the argumentation frameworks, and allow for outcomes through run-time semantic interpretation instead of being baked into the code. This last design builds an evolving (adaptable) agent as the organization evolves in its growth in response to how organizational roles, responsibilities, and workflows are evolving over time. 

Furthermore, by anchoring agents in organizational processes, tuning for autonomy, purpose to tools, and design flexibility through distributed architectures, organizations can start to deploy multi-agent systems which can be considered more than automation. They can evolve to become ‘living,’ responsive networks which operate in much the same way a team of people would work, engaging, collaborating, and communicating. based on past building blocks of input-output formats.

Selecting a Suitable Agent Architecture

The architecture of a multi-agent system significantly influences its versatility and durability. A blackboard architecture consists of a centralized place where agents register their part, and a controller manages activities. While this is a simple, managed architecture; simple architecture means connections might be fragile because they depend on a single point of control and cannot adapt easily to dynamic environments.

The Asynchronous Agent-Oriented System Architecture (AAOSA), on the other hand, utilizes a distributed model, where agents decide for themselves if they are able to respond to a request, where they gather requirements, find down-chain agents to delegate a request, and elevate unmet requirements back up the chain. The architecture is robust and scalable and fits very well in today’s organizations, fitting the need of functionality under adaptability and survivability. Distributed is much more aligned with how organizations function, i.e., the responsibility is shared across networks, rather than being a singular point of responsibility positioned at the top.

Training Agents to Operate as AAOSA Participants

For agents to effectively operate in AAOSA, they need to undergo training, and construct prompts to think in the same manner as any other participant in a digital ecosystem. If an agent receives a request, he/she should first consider their tools and capabilities. If there is an inability to completely respond at that level directly, agents should query down-chain agents for their requirements, synthesize results and delegate. They should also continually be prepared to act in the capacity of being a down-chain participant when called upon by another agent. By being an automatic participant in a cooperative action, the agents transcend just the production of outputs; they are now collaborating, negotiating, and responding with flexibility as equals joined in a community.

TLDR: Establishing the Foundation for Organizations to be AI-Driven 

Multi-agent systems are not simply stitching together technologies; they are the next evolution in organizations intelligence. At their core, multi-agent systems replicate the work dynamic in the way people collaborate, work together, and develop structure in ways of work processes in complex networks of roles, responsibilities, and work processes. By defining, understanding, and representing their collaborative dynamic, procedural processes, systems and connectives, leads to transformative value beyond automation.

Mirror Human Work: Multi-agent systems mirror work dynamics that would otherwise only be found in human teams and enhance one work type to provide, enhance, and optimize those processes digitally, thus retaining and providing human value. Taken together, agents are able to modify and command workflow tasks, communicate with each other in a similar fashion of how teams engage (collaborate), and together, support a human-based decision making process. 

Increase Productivity: Organizations can seamlessly offload most productive work (dependent, distributed, and high knowledge transfer) to agents and elevate professionals to a more productive (higher and more strategic value) and improve quality of output across the organization. 

Improve Resilience: Organizations are realizing new levels of operational resiliency as agent solutions will be more modularised and distributed. Consequently, if something fails (or has to be taken offline as part of a process update) the remainder of the operation will run unless this action is dependent on the micro-agent. 

Adapt to Change: The adaptive nature of multi-agent frameworks are fluid. As a result the organisation can inject new tools and evolve to third-party systems to remain relevant with the financial and social change. 

In short, multi-agent systems provide the precursor to AI-based organizations of the future—an uninterrupted organization that works seamlessly, spills, and scalable-two systems that are capable of constant learning, adapting, and growth. Organizations that begin to experiment with the multi-agent systems will be the applicable to thrive on the other token of the diversifying, digital/AI enabled world. 

Shape the Future of AI with AIU

At Atlantic International University (AIU), we believe that innovation in technology and organizational design starts with education. We demonstrate this belief through our philosophy of andragogy, omniology, and purposeful learning. Our education model engages students in growing areas imaginable, such as AI agents, multi-agent systems, and digital transformation, while allowing them to set the path for their own education towards personal and professional goals.