If you've heard about AI Agents and you're intrigued about them, this is the article to read.
AI agents are nothing new, since the start of the AI research, the goal has been to create artificial humans. They can interact with humans, perform tasks, and change their environment.
With the advent of LLMs, and their reasoning capabilities, there has been a influx of work to make autonomous and semi-autonomous AI "agents". These agents can work independently and interact with humans occasionally to get feedback and directions.
Primarily, the AI agent is not separate paradigm, but rather a different way of architecting AI solutions.
Think of a AI pipeline designed around researching a topic. The following steps are required.
Search the internet for the topic
Gather the links to relevant articles
Read the articles and collect findings
Summarize and organize the findings for human consumption.
A "traditional" AI system will be architected as a sequence of steps. Each step, produces the output for the next step and the final step completes the "chain".
Use Google Search API to get search results.
Store the links in a database or store.
Scrape the article links and collect the text.
Digest the text from each document and create a summary.
An agent-based architecture would also do the same, but it will be architected slightly differently. Firstly, we'll identify the tasks that need to be performed and the tools are resources needed.
Tasks needed to be performed
Search the internets for topic keyword.
Scrape text from web-documents
Store text in database or store.
Summarize text from each article.
Collect summaries from all articles and produce final article for human consumption.
Google Search API
Text Scraping Library
Now, let's map the tasks to different "agents", who will each have access to the tools.
Task description: You are a researcher who needs to collect article links around a given topic. You have access to google search API and can use it to gather the links. Once you can save the links in the database, using the database API.
Task Description: You are an analysts who needs to read an article around a topic and summarize it. Once you have created the summary you can store it using the database API.
Task Description: You are a writer whose job involves creating an article from multiple sources. You will be provided a list of summaries from various articles, and you need to write a article summarizing them all.
Nothing fundamentally has changed from the traditional AI pipeline, mind you, the process will still work in a similar fashion. What has changed is the "architecture" of the solution.
Each agent can be independently worked on by different teams / people and fine-tuned for a specific task.
Agents can be given different tools, for example, Bing Search API in addition to Google Search API.
This also enhances the reusability of the system, as each component or agent is modular and can be put in a different project or as a part of a different solution.
Frameworks like CrewAI and Microsoft Auto-Gen make it easier to create these agents, and setup interactions between them.
Agents share tools, to interact with the outside world.
Agents reason and solve limited problems within their task context.
Agents can interact with each other, and also humans.
Agents can supervise other agents and give feedback to improve the results.
So to summarize, Agents are a perspective shift rather than a paradigm shift and nothing revolutionary themselves. They offer better encapsulation of logic and allow for systems to make the AI pipelines more modular, and specialized.
Core Maitri is an enterprise software consultancy specializing in Excel-to-Web, AI Integration, Custom Software Programming, and Enterprise Application Development services. Our approach is deeply consultative, rooted in understanding problems at their core and then validating our solutions through iterative feedback.
If you're interested in rapid software development to propel profit and growth, contact us!