Original subtitles

- [Instructor] Let us take a step back

and understand how agent AI works.

Agent AI works using a system called agent AI workflow,

which is made of reasoning LLMs, tools,

processes to reflect and communicate

with other agents to automate complex tasks.

It all starts with the user prompting the agent

to automate a complex task.

The agentic system breaks the complex task

into simple tasks.

Each task is assigned to an agent.

The agent uses tools to get results.

The agent then reflects on its results to improve it.

Agents debate and communicate with other agents

to achieve the best results.

So to build an agent AI,

we need to start with a large language model,

and we will be using the reasoning power of the LLM.

So we need a LLM that is able to do reasoning.

The agent AI is deployed to users

and the user prompts a chatbot with an ask.

For example, you as a user could ask an agent AI,

such as an agent GPT,

to build out a marketing plan for a new product

you want to launch.

In this case, the agent AI takes the user prompt

and breaks it into simple tasks

and begins a step-by-step execution.

A LLM can chat with us by looking back into its learning

to get us answers for any prompt we make.

It does not have the ability to execute a computer program,

so the agent needs to use tools

to do runs to get results for the LLM/

For example, one step in the marketing launch plan

for your product could be to come up with a name

for the product.

The agent can use a simple web search as the tool

to find out if there are similar products out there

in the industry and come up with a unique name

for your product.

The agent uses processes to iterate and improve its finding.

In our example,

the LLM can reflect on the name it came up with

and ponder to improve the results.

It could check if the name it came up with

does not have a bad meaning in any global languages,

and check if the name is easy to pronounce

and to use as a hashtag in a campaign.

Multiple agents communicate with each other

till they arrive at the final solution.

Other agents might be working on other steps of the task

to come up with a promotion plan

or building out a logo for the new product.

Finally, the agent communicates with the user

with the full marketing plan for the product.

Sometimes the agent might be fully automated

without needing the first step of a user prompt.

In this case,

the agent could communicate with its environment.

Can you think of an agent that communicates

with its environment using sensors

instead of communicating with the user who prompted it?

What about an agent taking care of the safety of a granary?

The agent can adjust actuators to cool down the granary

to keep the safety levels by watching for weather change

and by tracking humidity levels in the granary.

The possibilities are endless.