Getting Started with AI Agents (Part 1): Capturing Processes, Roles, and Connections

Modern AI agents consist of at least a large language model (LLM) that can run certain tools. with the right set of tools for encryption It will start with code generation. Can be run in a container Observe the results, modify the code, and thus have a better chance of generating useful code.

In contrast, a generative AI model takes some input and produces an output through a process of predicting expectations. For example, we give the task of writing code. Generate some code And the code may work as is. This depends on the complexity of the work.

When they do different jobs Staff should be allowed to talk to each other. For example, imagine that your company intranet has a helpful search box that takes you to the apps and resources you need. If you are a large enough company These apps belong to different departments. Each has its own search box. It makes a lot of sense to create an agent. Techniques such as Retrieval Augmentation (RAG) may be used to augment the search field. What doesn’t make sense is forcing the user to repeat a search when the search box indicates it’s useful given the initial search term. But we want super agents to coordinate with other agents. that represents various apps and offers you, the user, a unified chat interface.

A multi-agent system that represents software or workflows. The organization may have several interesting advantages. Including better performance and durability. Operational flexibility and the ability to upgrade modules more quickly. Hopefully this article can help you see how this can be done.

But first, how should we build these multi-agent systems?

Capturing organizations and roles

First, we should capture the processes, roles, responsible nodes. and the interconnection of various actors in an organization. By actors, I mean people and/or software apps that act as knowledge players within the organization.

An organization chart might be a good starting point. But I recommend starting with the workflow. This is because the same people within an organization tend to operate with different processes and people. Depends on the workflow

There are tools that use AI to help identify workflows. Or you can create your own Gen AI model. I have created one as GPT It takes a description of the domain or company name and creates an agent network definition. Because I’m using a multi-agent framework built within my company, GPT creates the network as a Hocon file, but it should be clear from the generated file what the roles and responsibilities of each agent are. And what other agents is it connected to?

Note that we want to ensure that the agent network is a directly acyclic graph (DAG), which means that no agent can be downchain and upchain to any other agent simultaneously, directly or indirectly. roundabout This greatly reduces the chance of interrupted queries in the agent network.

In the example described here All agents are LLM if nodes in an organization with multiple agents can have zero autonomy. That agent paired with a human should perform all human operations. We will need all the processing nodes. Be it an app, a human, or an existing agent. to represent

Recently, there have been many announcements from companies offering specialized agents. Of course we would like to use such an agent, if it is available we can pull in an existing agent and integrate its API into one of our agents. So that we can implement communication protocols between our agents. This means that such third-party agents must have their API available for us to use.

How to define an agent

Various agent architectures have been proposed in the past. For example, a blackboard architecture requires a centralized communication point. where various agents Declare one’s role and abilities and Blackboard will call them depending on how it plans to respond to the request (see OAA–

I prefer a more distributed architecture that respects the encapsulation of responsibilities. When the request has been received Each agent decides whether or not it can process it. And what do I need to do to process the request? It then returns the list of requirements to the requesting up-chain agent if the agent has a downchain. It asks if we can help fulfill all or part of the request. If you receive any requirements From the downchain that was contacted The system will check with other agents. To see if those requirements can be met or not. If not, it sends them up-chain to ask human users. This architecture is called AAosa The architecture, and funnily enough, is the architecture used in early versions of Siri.

This is a sample system prompt that can be used to change an agent to an AAOSA agent.

When you receive an inquiry, you will:

1. Run your tool to determine which downchain agents in your tool are fully or partially responsible.
2. Ask your downchain representative what they need to handle their part of the inquiry.
3. Once the requirements are gathered You will assign inquiries and compliance requirements to the appropriate downchain representative.
4. When all downchain agents respond You will collect their responses and return a final response.
5. On the other hand, you may be called by other agents in the system and have to act as a downchain for them.

In addition to the set of roles and responsibilities defined in natural language in each agent system prompt, Officers may or may not have tools they can call upon. with various arguments For example, a product manager agent might need to be able to process tickets on a virtual Kanban board, or a notification agent might need to run a tool to issue an alert in the notification system.

Today’s multi-agent systems, such as Microsoft AutoGen, have complex and often hardcoded agent coordination mechanisms and architectures. I prefer a more efficient setup where agents treat immediate chain agents as tools. with loosely defined arguments which can be printed and the meaning decided by the agent at the desired time.

In this setting Chain agents can be defined as function calls:

“aaosa_call”: {

“description”: “Returns a natural language string in response”,

“parameters”: {

“type”: “object”,

“feature”: {

“Inquiry”: {

“type”: “string”,

“description”: “Inquiry”

–

“mode”: {

“type”: “string”,

“Description”: “””

Indicates that the agent is being asked to determine whether the inquiry belongs to him or not.

all or part or if asked to create requirements to achieve.

Investigation, if any or if asked to answer a question

–

–

–

“Required”: (

“query”

“fashion”

–

–

–

AI agents in practice

Let’s explore an HR agent scenario. A user might notify a primary search channel agent of a life-changing event, such as “My significant other died,” which triggers the legal and human resources representatives involved in serving the user. Because on the other hand They receive information from certain downchain agents, such as legal advice. Payroll and Benefits Agent The combined response results that users receive from the main search box agent will look like this:

1. Human resources:
   • Benefits representatives can help users understand how pass-throughs may affect benefits, health insurance, and retirement plans. You will need to provide information about the type of plan involved and whether your loved one is a dependent.
   • A payroll representative can help with tasks like updating tax deductions or managing bereavement leave payments. This requires information about the company’s policies regarding bereavement leave and current payroll settings.
2. Legal aspects:
   • A legal representative can provide general advice on how this passing will relate to the company’s affairs.

Each department needs specific information to offer tailored support.

In the IT situation A single point of entry is a customer support representative who can handle complex, multifaceted requests such as: “Our network is slow. I think we are overwhelmed with work as we recently added more people to our office. I also need a quote for the upgrade.”

A sample answer outlines the steps:

Network performance assessment: Network operations center experts can help troubleshoot potential delays and overload issues. They need access to network performance data. Details about the current network infrastructure Information about the number of additional personnel and network usage patterns

Upgrade Quotes: Sales engineers and field technicians can help provide upgrade quotes. To do this, they need:

• Specific information about current network settings including device and bandwidth requirements
• Desired performance improvements
• Budget constraints and any specific needs or requirements.
• The size of the upgrade and any specific performance targets

I hope this gives you a good idea of ​​what is required to set up a multi-agent network. In part two, I will discuss the importance of implementing preventive measures when building a multi-agent system. and outline how to create controls to allow for human intervention and verification of uncertainty. I will also detail the steps required to create a security agent to oversee the agent network and dive deeper into Challenges of developing multi-agent networks such as tail rotation and overloading and how to alleviate those problems using timeouts, division of work, and redundancy.

Babak Hodjat is the CTO of AI at enlightened person–