Starting with AI agents: Capturing roles, processes and connections
Join our daily and weekly emails to receive the latest news and exclusive content about AI. Learn More
An AI agent of today is at least a large-language model (LLM), which has been given the ability to call certain tools. A generative AI model, on the other hand, takes inputs and predicts expectations to produce an output. It can be given a coding job, which it will complete. Depending on how complex the task is, the code it generates may be useful. Imagine your intranet, with its search box that directs you to apps and resources. If your company is large enough, you may have separate search boxes for each of these apps. To enhance the search box, it makes sense to use techniques such as retrieval-augmented generation (RAG). It is not logical to make the user repeat the query after the search box has already identified the query as relevant. We would rather have the top agent coordinate with the other agents who represent various apps to present you, the user, a consolidated chat interface.
An agent system that represents software or an organisation’s workflows has several advantages. These include improved productivity, robustness, operational reliability, and the ability to upgrade different modules faster. Hopefully, this article will help you see how this is achieved.
But first, how should we go about building these multi-agent systems?
Capturing the organization and roles
First we should capture the processes, roles, responsible nodes and connections of various actors in the organization. By actors, I mean individuals and/or software apps that act as knowledge workers within the organization.
An organizational chart might be a good place to start, but I would suggest starting with workflows, as the same people within an organization tend to act with different processes and people depending on workflows. There are tools available that use AI for identifying workflows. You can also build your own Gen AI model. I built a GPT that takes a description of a company or domain name and generates an agent network definition. The GPT generates the network in Hocon files because I am using a multiagent framework that was built at my company. However, it should be obvious from the generated files which agents are connected to each other and what their roles and responsibilities are. It means that agents cannot be down-chained and up-chained to each other, either directly or indirectly. In the examples given here, the agents are all LLM-based. If a multi-agent node can have zero autonomy then the agent should be paired with a human counterpart to run all operations. We will need all processing nodes, be they apps, humans or existing agents, to be represented as agents.
Lately there have been many announcements by companies offering specialized agents. If such agents are available, we would certainly want to use them. We can use an existing agent by wrapping its API in one of our agents. This will allow us to make use of the inter-agent communications protocols. This means that such third-party agents will need to have their API available for us to use.
How to define agents
Various agent architectures have been proposed in the past. A blackboard architecture, for example, requires a central point of communication, where agents declare their capabilities and roles, and then the blackboard contacts them based on the way it intends to fulfill a particular request (see OAA). After receiving a request from an up-chain agent, each agent decides if it can handle it and what steps it needs to take to fulfill the request. The agent will ask its down-chains if they are able to help with the request. It checks to see if other agents can help fulfill any requests from down-chains. If not, then it sends the agent up-chain for the user to ask. Here is an example system prompt to help turn an agent into an AAOSA Agent. You will need to do the following when you receive an inquiry:
Call down-chain tools in order to find out which agents are responsible for it.
Ask them what they require to deal with their part of inquiry. You may be called upon by other agents to perform a similar function. For instance, a product manager agent may need to be able to process various tickets on a virtual Kanban board, or an alerts agent may need to call a tool to issue alerts in an alerting system.
Current multi-agent systems such as Microsoft AutoGen have elaborate and often hardcoded agent coordination mechanisms and architectures. I prefer a more robust setup where agents treat their immediate down-chain agents as tools, with loosely defined arguments that can be typed, and the semantics decided by the agents at the time of need.
In this setup, a down-chain agent can be defined as a function call:
“aaosa_call”: {
“description”: “Depending on the mode, returns a natural language string in response.”,
“parameters”: {
“type”: “object”,
“properties”: {
- “inquiry”: {
- “type”: “string”,
- “description”: “The inquiry”
- },
“mode”: Or if the agent is asked to specify requirements for fulfilling
the query, if applicable. {or if it is being asked to respond to the inquiry.
“””
},
},
“required”:
}
}
AI agents in practice
An agent network for a company’s intranet.|Or if the company is asked to provide a response to an inquiry.}
Let’s look at a HR agent scenario. The main search agent might be notified by a user of an important life event such as, “my significant partner passed away.” This would trigger the legal and HR agents to assist the user. It is because these agents received the information from their agents further down in the chain, such as legal advice, payroll agents and benefits agents. The user will receive a consolidated answer from the agent in the main search box. It will look something like this:
HR:
You’ll need to provide information about the types of plans involved and whether your significant other was a dependent. You’ll need to provide information about the types of plans involved and whether your significant other was a dependent.
The payroll agent can help manage tasks such as updating tax deductions or handling bereavement leave pay. It is necessary to know the company’s policies regarding bereavement leaves and how payroll is currently set up.
Each department needs specific information in order to provide tailored support.
In a scenario involving IT, the agent who can deal with complex, multi-faceted questions, such as “Our network runs slowly”, is the one to contact. I believe it’s overloaded because we added staff to our office recently. They will need to know the network performance data, details about the current infrastructure and information on the number of personnel added as well as their usage patterns. To do this, they need:
Specific information about the current setup, including equipment and bandwidth requirements.
Desired performance improvements.
Budgetary constraints and any specific preferences or requirements.
The scale of the upgrade and any specific performance goals.
I hope that you now have a good idea about what is required to set up a multi-agent network. To do this they need:
Specific information about the current network setup, including equipment and bandwidth requirements.
Desired performance improvements.
Budgetary constraints and any specific preferences or requirements.
The scale of the upgrade and any specific performance goals.
I hope this gave you a good idea of what is required to set up a multi-agent network. In the next installment, I’ll discuss the importance and how to implement safeguards for multi-agent system creation. Also, I’ll outline the controls that can be built in to allow human intervention and uncertainty checking. Welcome to the VentureBeat community! [
“inquiry”,
“mode”
]
DataDecisionMakers
Welcome to the VentureBeat community!
DataDecisionMakers is where experts, including technical people doing data work, can share data-related insights and innovation.
DataDecisionMakers
Welcome to the VentureBeat community!
- DataDecisionMakers is where experts, including the technical people doing data work, can share data-related insights and innovation.
- If you want to read about cutting-edge ideas and up-to-date information, best practices, and the future of data and data tech, join us at DataDecisionMakers.
- You might even consider contributing an article of your own!
- Read More From DataDecisionMakers
