Chatbots and Agents
A chatbot is a computer program that allows users to converse or question the application and receive responses that give the impression of a conversation. Initially they relied upon basic pattern matching and delivering canned script responses. Basically they would analyze the input statement from the user, identify keywords that would match pre-defined answers and respond with that data.
Some early techniques like “Mirroring” from ELIZA developed at MIT Chabots would respond to the user by repeating what they inputted in a form of a question giving the impression that it was listening to them.
These early chatbots had a "Teacher Mode" and some basic learning. Using the plus sign (+) the user could add new rules via keywords. They were limited in answers to responding with small scripts
As technology advanced so did chatbots like the 1995 rollout of A.L.I.C.E. (Artificial Linguistic Internet Computer Entity) leveraged a massive library of patterns. AOL rolled out the SmarterChild in 2001 on its popular messenger platform that provides weather, movie times, news, etc.
With the move to mobile platforms voice-activated AI assistants like Siri, Alexa and Google Assistant used advances in Natural Language Processing to better understand user requests and provide ability to execute the task requests.
The release of ChatGPT in 2022 allowed users to interact with LLM’s or Large Language Models that could write computer code, generate images, create music and solve complex problems.
The AI Agent
While Chat allows users to interact with AI the concept of an AI Agent is the process that allows for the completion of AI related tasks.
Agents create plans and break down task completion into a series of steps.
They have memory to keep track of what was tried in order to prevent ending up in an endless loop.
Agents can use RAG (Retrieval-Augmented Generation) to retrieve external database information and leverage information from previous sessions.
An important concept of Agents is that they can use tools such as web search, code interpreters, database connectors to complete tasks.
They have feedback loop ability to look at the result and if an error occurs the agent can move to a different approach until the goal is met.
While a chatbot follows instructions like summarizing this report an agent can use tools (Browsers, SQL, API’s) to research companies and update systems. Agents can figure out how to complete tasks on their own. on its own.
AI Agents use Large Language Models (LLMs) and act as a brain by creating step-by-step plans. They follow a ReAct (Reason + Act) pattern where they observe the goal, break it into smaller tasks, use tools for task completion and review results. The Agent will repeat the process until the goal is complete.
Google AI Agent Works
A typical Google agent has four core capabilities.
Reasoning & Planning: They will use models like Gemini to break down complex requests into smaller manageable tasks. The agent build its own plan and revise it when needed to achieve goals.
Memory: The Agent can remember previous steps and can learn from what worked or what didn’t.
Tool Use: Uses Tools to access databases and API’s and can update systems. Agents can use functions to update databases and systems connecting to internal applications like calendars and email.
Autonomy: It doesn’t need the user to provide step by step instructions. You give it a goal and it figures out the steps to achieve it.
Agents are not limited to leveraging LLM’s, they can also use standard coded solutions in task execution.
Multi-Agent Systems can be used to develop complete solutions where each agent can handle specific functions. For example you can build an AI system for stock trading. One agent can report performance, another can research stocks while a third agent can execute trades. Each can communicate and interact with each other.
In this scenario each agent performs a specific role and each has goals to achieve. They are designed to specialize in specific areas of monitoring, researching and trade execution. Also, they are designed to communicate and collaborate with each other in order to help achieve an ultimate goal of producing results that involve maximizing the amount of money that can be made.
Agentic AI and Agentic Workflows
Now that we have an understanding of AI Agents we need to see how to implement them. We can use the Agentic AI process where the agent's reasoning determines the execution path to achieve its goal. or we can implement an Agentic Workflow where the developer can pre-designs the way AI Agents interact.
AI Agent
Has a defined role to interact with LLM’s and has various tools it can use to interact with systems. It is the functional implementation of a specific AI goal. It combines the defined role with sets of tools like APIs and system integrations to deliver results. AI Agents are the entities that execute clearly defined instructions and use tools to interact with external systems.
Agentic AI
Has reasoning capabilities meaning it can think through the request logically before implementing a solution. It has the capability and intelligence to reason, plan and act without the need for interaction but can allow for human-in-loop to validate results. It leverages the LLM’s ability to think in loops to develop logical plans, detect errors and correct issues. When it finds an error it can reason through why it happened, and adapt the process to resolve the issue.
Agentic AI Workflows
Allows the developers to define a structure for AI solutions to follow. They define a sequence of steps to coordinate with the AI Agents to guide them to achieve the specific goals. It will enforce a logical order such as searching, critiquing, and then summarizing and act as a guardrails to ensure that AI is implemented in the way the developer has designed the solution.
The transition from AI Chatbots to Agentic systems is a change in how to approach problem solving.
While chatbots can get answers from specific prompts, Agents have tools to add meetings to calendars, retrieve and update data from database systems and interact with external software packages the corporation may use like CRM. Agents can plan, execute, detect errors, and correct those errors until the goal is achieved.
Agentic systems require a specialized structure, Reasoning processes logic and determines the next step. Memory to access current conversations and short term history. Long term knowledge retrieval via RAG or Retrieval-Augmented Generation. The Retrieval process allows an AI model access to data outside of its training set. It can search company sources like documents, databases, or specific news feeds for relevant information. The Augmentation process adds the information as context. The Generation process uses that context to create a more accurate, up to date and factual response.
RAG can reduce hallucination by providing actual facts rather than AI predictions. The AI sticks to the provided text rather than guessing. Use fresh data that was not used when the model was trained. Can access private data that the model does not have access to.
Agentic AI performs well at non-deterministic, complex, and open-ended tasks by providing paths that a developer may not think of. Agentic Workflows allow the developer to solve using a deterministic structure. When the steps are clear and should be executed in a certain way. Provides predictable solutions that can meet business and security standards.
In summary, all three elements are important when building agentic systems. Agentic AI possesses the reasoning power knowing how to use tools and iterating through ways to achieve the goals. AI Agents understand how to use tools and follow plans laid out. Agentic AI can instill rules in defining steps that the Agents are executed.
No comments:
Post a Comment