Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production — Transcript

Full transcript

1. 0:00Hi, let me guess.
2. 0:02The world around you is abuzz with agentic  AI systems and their massive potential.
3. 0:07So you decide to go off and  build an agentic system.
4. 0:10You look for the best  available framework out there.
5. 0:13And now, all of a sudden,  you have 17 GitHub tabs open,
6. 0:17five medium blocks bookmarked, and you  are still clueless on how to proceed.
7. 0:22Yes, we've all been there.
8. 0:24LangChain, LangGraph, Crew AI, AutoGen, Semantic Kernel.
9. 0:29There are so many powerful frameworks  out there, but which one would you pick?
10. 0:34And in order to answer that question,
11. 0:36you first need to understand which type  of agentic AI system you want to build.
12. 0:41In this video, we are going to cover five types  of most common agent AI projects or systems.
13. 0:47We are going discuss an example for  each of those and also list frameworks
14. 0:52which are best suited for those types  of agentic AI systems and projects.
15. 0:57First, let's get some basics out of the way.
16. 0:59What exactly is an agentic AI framework?
17. 1:02An agentic framework is a toolkit  for building agentic AI systems.
18. 1:06Let's understand with an example.
19. 1:08Let's say you have an agent  to analyze some sales data.
20. 1:12This agent goes to a database  and extracts the relevant data.
21. 1:16It analyzes the data, maybe  runs some calculations,
22. 1:19and then generates a report and crafts a response  that can be sent to the concerned person.
23. 1:24Now, there's a lot going on here, and  there's lot of coordination that's needed.
24. 1:28And it gets even more complex when  you have multiple agents working.
25. 1:31And that's exactly why you  need an agentic AI framework.
26. 1:35The framework is like a building block.
27. 1:37Unlike a chatbot application where you  just ask a question and get an answer,
28. 1:41the agentic system actually does a lot  of planning, acting, and iterating.
29. 1:53It is because of this complexity involved  that we need an agentic AI framework.
30. 1:57These are like building blocks that help  us deploy and manage agentic AI systems.
31. 2:04Now these have some predefined functions that help us
32. 2:08build agentic AI systems with more ease and  agility, such as we have predefined architectures.
33. 2:19We might have integration and monitoring tools.
34. 2:30We might also have some task  management capabilities.
35. 2:37And communication protocols.
36. 2:42Together, these features and functionalities  give agentic AI frameworks the capability
37. 2:47to allow us to build these systems with ease.
38. 2:50Like we discussed before, there are multiple  agentic frameworks available out there,
39. 2:55but they're not all competing  to do the same type of tasks.
40. 2:58In fact, they're optimized for  different types of agentic AI systems.
41. 3:02Most agentic AI systems and projects we are working  on today fall into one of the five categories.
42. 3:07First, we have linear workflows.
43. 3:09We have autonomous AI agents or  autonomous multi-agent systems.
44. 3:14We have role-based AI systems,  production orchestration systems,
45. 3:18and then we have rapid prototyping.
46. 3:21Let's dive in and understand  each of these in more detail.
47. 3:24Let's start with the simplest  one, linear workflows.
48. 3:30Now, this type of an agentic AI system is where  things progress in a step-by-step fashion.
49. 3:35It is more predictable what's gonna happen next.
50. 3:41And the steps follow a certain sequence.
51. 3:47For example, consider a customer-facing  application, let's say a customer support agent.
52. 3:53The role of this agent is when a user asks a  question, the agent is going to take the question
53. 3:58and search the knowledge  base for relevant responses.
54. 4:01It's then going to craft a response  and send it back to the user
55. 4:04and maybe take an additional action  such as creating a support ticket.
56. 4:09Now, if you observe, these steps are progressing  in a certain fashion, in a certainly sequence,
57. 4:14and these systems are more useful when  you need the flows to be more reliable.
58. 4:22There isn't a need for multiple agents  to collaborate to make this happen.
59. 4:26And that gives you more  control on how things progress.
60. 4:33A good example of frameworks that are suitable  for this kind of a setup include LangChain.
61. 4:42And LlamaIndex.
62. 4:47LangChain is more suited for setups where multiple  steps need to be happening in a certain sequence.
63. 4:54LlamaIndex is highly suitable for
64. 4:57heavy applications that are heavy  on the data retrieval and indexing.
65. 5:03For more complex setups, you could also  use LangGraph, which is also by LangChain.
66. 5:12Next, we have the autonomous agentic AI systems.
67. 5:17In these systems, you typically give AI a goal.
68. 5:23And have it figure out how to accomplish it.
69. 5:26So in this system, it's very common to see  multiple agents collaborating together.
70. 5:36These agents talk to each other  to accomplish the common goal.
71. 5:40A good example of this could  be an AI coding assistant.
72. 5:43You could have a planner agent  that plans the solution for you.
73. 5:47You could a coder agent that  actually writes the code for you,
74. 5:51and a reviewer agent, that is reviewing the code,
75. 5:53making recommendations, and  also helping with the debugging.
76. 5:56These agents are constantly talking to each other  in order to give the best code possible to you.
77. 6:02So, in this kind of setup, the  problem is usually open ended.
78. 6:07And that's the kind of problems this  kind of setup is most helpful for.
79. 6:14So, frameworks that work best for this  kind of a scenario include AutoGen.
80. 6:23You could also use experimental  setups like Baby AGI.
81. 6:29And CrewAI could also
82. 6:32be helpful for designing these kind of  systems where problems are open-ended
83. 6:38and multiple agents are collaborating  together to achieve a shared goal.
84. 6:43Next, we have the role-based agentic AI systems.
85. 6:47These are kind of similar to  the autonomous agentic systems
86. 6:50where there are multiple agents collaborating.
87. 6:53So it is also a multi-agent setup.
88. 7:01But what makes it different is that each  agent within the setup has a defined role.
89. 7:10They are still communicating with  each other to accomplish that goal,
90. 7:14but they are operating within the confines or  the constraints posed by their role descriptions.
91. 7:20They are working together,  but with clear boundaries.
92. 7:23A good example of this could  be a content generation agent.
93. 7:27Here, you could have a researcher  agent that goes on the web and fetches
94. 7:31all the material that's needed  to write a piece of content.
95. 7:34There could be writer agent  that looks at all the content
96. 7:37that has been fetched and writes up an article.
97. 7:39That goes out on a social  media website, let's say.
98. 7:42And then there could be an editor  agent that's looking at the article
99. 7:46that has been written and make some edits to it.
100. 7:49Now, they have very clearly  defined roles and they don't
101. 7:52go into other agents roles when they do this.
102. 7:55They have discussions, but  they're strictly confined
103. 7:58to the description that has given  to them for their particular roles.
104. 8:01A good framework that is  applicable here is CrewAI.
105. 8:09But you could also use AutoGen  with some structures around it
106. 8:14for this kind of an agentic AI systems.
107. 8:16There are also some niche frameworks that  are applicable to very specific tasks.
108. 8:20Like, for example, for software development  kind of tasks, you have ChatDev.
109. 8:28So these are the kind of frameworks that you  would use for role-based agentic AI systems.
110. 8:34Next up, we have the production  orchestration systems.
111. 8:38Like the name suggests, this is when AI  moves out of the experimentation phase
112. 8:43and gets real or moves into a real-world system.
113. 8:49These kind of systems require  deep integration with APIs,
114. 8:57databases,
115. 9:01and business workflows.
116. 9:08Consider the example of an AI operations agent.
117. 9:12This agent detects alerts, searches  the documentation for the alerts,
118. 9:17and then runs some automation scripts
119. 9:19and sends summaries in a real world  scenario within an organization.
120. 9:22Good examples of frameworks that are  suitable for this kind of an AI system,
121. 9:27include agent framework,
122. 9:33which essentially is a combination  of semantic kernel and autogen.
123. 9:37Another good example here is LangGraph.
124. 9:45Which works for well-structured,  multilayered applications.
125. 9:50Ancient framework is suitable both for  orchestration as well as for running
126. 9:55autonomous workflows.
127. 9:57Last but not least, you  have the rapid prototyping.
128. 10:02So you always don't need a perfect architecture.
129. 10:05You just need to check if  your idea would work or not.
130. 10:09These types of systems are best when  you need to quickly validate ideas.
131. 10:16It helps you build quick prototypes.
132. 10:21To see if you can bring your ideas to reality.
133. 10:25These kind of systems are where you ideally have  a user interface where you can drag components
134. 10:31and bring them onto a canvas to  build quick workflows and test ideas.
135. 10:36Examples of tools or frameworks that  are useful here include LangFlow.
136. 10:45And Flowise.
137. 10:49These tools offer you a good graphical
138. 10:51user interface where you can  bring components onto a canvas
139. 10:55and connect models and workflows  and quickly test out your ideas
140. 10:59that you can later on take into production.
141. 11:01These are very quick for rapid  prototyping and hence the name.
142. 11:05So when choosing a framework, do  not ask which framework is the best.
143. 11:09Instead ask what kind of  system am I trying to build?
144. 11:12If it's predictable, use a workflow approach.
145. 11:15If it is exploratory, use the autonomous agents.
146. 11:18If it needs teamwork, use role-based systems.
147. 11:20If it's going into production, use the  production orchestration frameworks.
148. 11:24And if you are testing ideas,  use the rapid prototyping tools.
149. 11:28Because the right framework depends  on whether you're building a pipeline,
150. 11:31a team of agents, or a production AI system.
151. 11:34Which agentic AI framework is your typical go-to?
152. 11:37Feel free to comment below and  don't forget to like and subscribe.
153. 11:41Thank you.