Agents 101 for AI Engine 🤖

Overview

Welcome to Agents 101 for AI engine! This course is designed to introduce you to the overall system of Fetch.ai, and shows how to build agents to be accessible to the AI Engine and integrate with other projects. This course is a parallel track to Agents 101 🤖↗️ which instead focuses only on Agents creation.

At any stage, if you encounter issues or have questions about specific terms or topics, our support team is available on Discord ↗️ (opens in a new tab) to assist you.

Quick overview of Fetch.ai

Fetch.ai is developing a platform to help build an AI-enabled decentralized digital economy. Agents are programs that can make choices on their own for individuals, companies, and devices. Agents are the actors and the heart of Fetch.ai Ecosystem.

To build your solution on Fetch.ai, you need to understand four concepts: Agents ↗️, Agentverse ↗️, AI Engine ↗️, and DeltaV ↗️.

Agent, Agentverse, AI Engine and DeltaV

Let's try and keep this simple 🎉:

• Agents: These are programs that work as you instruct them. Classically, they are part of a multi-agent system where they have a task and communicate with other agents. But their primary use in this case is to represent a service, a function or data that can be accessed by DeltaV.

• Agentverse: It is a managed hosting platform, an online IDE, and an entry point to AI Engine for Agents. You would register your Agents on Agentverse to be a service in the network.

• AI Engine: The AI Engine is a system that combines Agents with human-readable text input to create a scalable AI infrastructure that supports Large Language Models (LLMs). It's an AI system that can convert human text into actionable (computer) commands.

• DeltaV: DeltaV is an assistive chat interface that allows human input to interact with agents registered through the Agentverse.

  ⚠️

  Agents are programs, whereas services are the functions your agents offer.

In a nutshell, AI Engine indexes Agents, and translates human input into tasks. It may sound like a search engine, but it does much more.

The AI Engine allows chaining. Say that you ask DeltaV "Tell me about the animal Zebra". The AI Engine would then look for an agent that could give you a summary, as well as another agent that could provide information about zebras. The AI Engine understands that the Summary Agent requires text to parse and summarise. So it would contact the Zebra Agent and get the information from it. The AI Engine would then call the Summary Agent with the Zebra Agent information so that the former would return the summarized text.

It works like this:

Simply put, when the AI Engine selects a response, the subtasks available to be called for that task are already known to it.

Of course, this could be more complicated. What if every sensor on a car was an agent? You could ask your car for diagnostics on a system, and each component in that system would independently give its status - making everything extremely complex. That's where the power of the AI Engine comes in. It understands the schema of each Agent, so there would be little need for each component to be compatible in communication with each other.

Okay, so how does this all work?

To understand this, we need to look into a few more concepts:

Here, two separate processes are happening. The first part of the sequence diagram registers two agents:

1. Firstly, agents register to a smart contract thus effectively paying to be discoverable.
2. Agents then register as a service on Agentverse (you'd usually do this manually).

Now that the Agents are registered and discoverable - let's move on on to the second part of the diagram. The second part of the diagram introduces the flow of the AI Engine. This shows the role of a user who has entered a search term into DeltaV.

1. A query comes in and it is sent to the AI Engine.
2. The AI Engine has an internal loop that updates services registered and indexed.
3. It also has another internal loop that creates embeddings of these services.
4. Upon receiving text, the AI Engine understands and maps the text to Services.
5. It now requests addresses from Almanac ↗️ to message these Agents.
6. The AI Engine sees that Agent (A) needs some text to perform its services.
7. It finds subtasks that can provide the required data.
8. It then contacts the subtask, and the required data is returned.
9. Data is passed to Agent (A) which returns a summary of the data.
10. The response is then returned to the user on DeltaV.

Tasks and subtasks are agent functions. In the context of the AI Engine, these are (sub)tasks, but you're developing Agents.

Okay, let's build a couple of Agents!

We've got a high-level idea of how this works. Let's build this on Agentverse so that we can get this running on DeltaV and see the whole stack connected.

If you've never installed anything for uAgent development, checkout these guides:

• Setup the development environment ↗️
• Install uAgents locally ↗️

Create your first agent: the text sentiment agent

We're going to run our agents locally, and set up a Mailbox on Agentverse.

In this guide, our first agent takes any text and returns a sentiment of contents with openAI.

We need to create a Python file for this example. We can do this by running: touch sentiment_agent.py

from uagents import Agent, Context, Protocol, Model
from ai_engine import UAgentResponse, UAgentResponseType
import openai
 
class SentimentResponse(Model):
    text: str
 
SEED_PHRASE = ""
AGENT_MAILBOX_KEY = ""
 
sentimentAgent = Agent(
    name="SentimentAgent", #or any name
    seed=SEED_PHRASE,
    mailbox=f"{AGENT_MAILBOX_KEY}@https://agentverse.ai",
)
 
print(sentimentAgent.address)
 
content_protocol = Protocol("text sentiment analysis")
 
def get_sentiment(text):
    response = openai.chat.completions.create(
 
        messages=[{
            "role": "user",
            "content": f"Sentiment analysis of the following text:\n{text}\n",
        }], model = ("gpt-3.5-turbo"))
 
    return response.choices[0].message.content
 
@content_protocol.on_message(model=SentimentResponse, replies={UAgentResponse})
async def sentiment(ctx: Context, sender: str, msg: SentimentResponse):
 
    sentiment = get_sentiment(msg.text)
    await ctx.send(
        sender, UAgentResponse(message=sentiment, type=UAgentResponseType.FINAL)
    )
 
sentimentAgent.include(content_protocol, publish_manifest=True)
sentimentAgent.run()

Okay, we've got the whole code block above for our first agent. This agent does not run on Agentverse currently, as openai is unsupported.

There are a lot of really cool, neat things you need to know from code above. Let's explore it:

class SentimentResponse(Model):
    text: str
 
SEED_PHRASE = ""
AGENT_MAILBOX_KEY = ""
 
sentimentAgent = Agent(
    name="SentimentAgent", #or any name
    seed=SEED_PHRASE,
    mailbox=f"{AGENT_MAILBOX_KEY}@https://agentverse.ai",
)
 
print(sentimentAgent.address)
 
content_protocol = Protocol("text sentiment analysis")

• The class SentimentResponse defines the structure of the message the agent expects to receive. In this case, each message should have text variable

• SEED_PHRASE is our agents unique seed.

• AGENT_MAILBOX_KEY mailbox key is our ID of our mailbox hosted on Agentverse.

• OPENAI_API_KEY is our API key to OpenAI's APIs.

We can now initialise the Agent and define the Protocol ↗️. To read more about Agent objects, please checkout the reference docs ↗️.

Then, we have the on_message() handler:

@content_protocol.on_message(model=SentimentResponse, replies={UAgentResponse})
async def sentiment(ctx: Context, sender: str, msg: SentimentResponse):
 
    sentiment = get_sentiment(msg.text)
    await ctx.send(
        sender, UAgentResponse(message=sentiment, type=UAgentResponseType.FINAL)
    )
 

We need a way for our Agent to receive messages. We do this by creating a function for our agent to handle all incoming messages from other Agents. We use on_message() decorator which activates the sentiment() function once our Agent receives a message matching the SentimentResponse Message data Model we previously defined. Internally you can see that sentiment() calls get_sentiment(), this is useful for you as a developer as this shows that calling another function inside a decorated function is easy to do.

We can finally run the Agent at the bottom of the script:

sentimentAgent.include(content_protocol, publish_manifest=True)
sentimentAgent.run()

Including the protocol in the Agent effectively allows other Agents to know how this Agent expects communication.

You can run this with: poetry run python sentiment_agent.py.

You'll see something like agent1qtc5m8xudkm6gjv98k3kxl3ydzhkhzsfx6pgla0mzlgsawlqkzkvcjh5en2 in the terminal.

Let's use this address to create a Mailbox on the Agentverse.

Creating a Mailbox for Sentiment Agent

A mailbox is a service provided by Agentverse that allows agents to receive messages when they're not running online. uagents library will poll for new messages (and old unread ones) when your agent is running. Pretty handy.

Let's go to Agentverse ↗️ (opens in a new tab) and create a Mailbox.

Add a name:

Let's paste in the Agent's address we printed before to screen:

Then, let's copy the new Mailbox API Key.

Update the AGENT_MAILBOX_KEY = "AGENT_MAILBOX_KEY" inline.

Now, you can restart your agent and it will be registered to the Mailbox.

Your terminal output will look similar to the following:

Poetry run python html_summary_agent.py
agent1XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX4
INFO:     [SentimentAgent]: Manifest published successfully: Text Sentiment
INFO:     [SentimentAgent]: Almanac registration is up to date!
INFO:     [SentimentAgent]: Connecting to mailbox server at agentverse.ai
INFO:     [SentimentAgent]: Mailbox access token acquired

Let's run this agent on deltaV, it'll be good to test, right?

We're going to need to create a service on agentverse, to get a comprehensive overview of what services are visit our agent services guide ↗️. A Service refers to an Agent enrolled within the Agentverse, and made discoverable by other agents and users for interaction. For instance, a service could be an Agent helping in booking an airplane ticket to some destination. Successfully registered Agent Services are discoverable and accessible through DeltaV and users can engage with them using natural language queries.

Let's go to the Agentverse ↗️ (opens in a new tab) and create a Service. You will probably need to create a service group to. A Service Group is just a name given to a group of services so you can link all your services neatly within AI Engine.

The following is the service for the SentimentAgent:

Really simple just defining what this function does. Simplicity for single function descriptions is good.

Let's head over to DeltaV and communicate with our agent.

I'm going to go ahead and ask:

"Provide the sentiment from this text "Petrol prices across the UK have hit an average 150.1p a litre, reaching the highest level since last November, according to the AA motoring group. Data also showed that average diesel prices rose to 158.3p a litre on Monday. Experts blamed a rise in oil prices because of tensions in the Middle East, as well as a weaker pound compared to the dollar. The AA predicted drivers would "re-tighten their belts on other spending". While overall inflation - which measures the pace of price rises - has eased, the cost of petrol and diesel both rose in March. "Inflation has been heading downwards at quite some speed but petrol's rebound to 150p a litre leaves a big boulder in the road," said Luke Bosdet, the AA's spokesman on pump prices. Prices are based on data collected by the Competition and Markets Authority (CMA) from nearly 3,000 forecourts and analysed by market research firm Fuel Prices Online. Data showed that 28.1% of the forecourts were charging at least 150p per litre for petrol on Monday, up from 23.8% a week earlier. But Mr Bosdet predicted that fuel prices are unlikely to be elevated for long with wholesale costs easing a little in recent days, adding that filling up "may not get much worse in the short-term". Oil prices have fallen after they spiked in the wake of Israel's retaliatory attack on Iran last week. Caroline Bain, chief commodities economist at Capital Economics, said: "Given that there will be some lag between spot crude oil prices in financial markets and prices on the forecourt, the latest rise is perhaps not so surprising." While Brent crude - a key benchmark for oil prices internationally - has receded from last week, it is still trading 13% higher at $87 per barrel compared to the start of the year."

Oil is often an issue in UK news. Let's find out the sentiment.

Note here, the service group is the name you defined when you created yours.

DeltaV tells us AI Engine has found a function for our query, and checks we want to send the information we defined.

Then, DeltaV get the response from our service and returns this to the user.

Your Agent running locally to you should have a similar output to this:

INFO:     [sentiment]: Manifest published successfully: text sentiment analysis
INFO:     [sentiment]: Almanac registration is up to date!
INFO:     [sentiment]: Connecting to mailbox server at agentverse.ai
INFO:     [sentiment]: Mailbox access token acquired
INFO:httpx:HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"

Okay, take a quick break then let's create the second Agent; it's time to introduce multi-agent workflows ☕.

Create the Webpage Summary Agent

The next agent in the chain is the Summary agent. Let's name the script web_summary_agent.py.

This Agent has the same structure as the sentiment Agent we previously defined.

Let's take a look.

    from langchain.chains.summarize import load_summarize_chain
from langchain_community.document_loaders import UnstructuredHTMLLoader
from langchain_openai import ChatOpenAI
from uagents import Agent, Context, Protocol, Model
from ai_engine import UAgentResponse, UAgentResponseType
import requests
import os
 
class SummaryRequest(Model):
    url: str
 
 
SEED_PHRASE = ""
AGENT_MAILBOX_KEY = ""
OPENAI_API_KEY = ""
 
summaryAgent = Agent(
    name="SummaryAgent",
    seed=SEED_PHRASE,
    mailbox=f"{AGENT_MAILBOX_KEY}@https://agentverse.ai",
)
 
summary_protocol = Protocol("Text Summariser")
 
print(summaryAgent.address)
print(OPENAI_API_KEY)
 
 
@summary_protocol.on_message(model=SummaryRequest, replies={UAgentResponse})
async def summarise(ctx: Context, sender: str, msg: SummaryRequest):
 
r = requests.get(msg.url)
with open("./temp.html", "w", encoding="utf-8") as f:
    f.write(r.text)
 
# Step 1: Initialize WebBaseLoader with the given URL
loader = UnstructuredHTMLLoader("./temp.html")
 
# Step 2: Load the document
docs = loader.load()
 
# Step 3: Load summarization chain
llm = ChatOpenAI(openai_api_key=OPENAI_API_KEY, temperature=0, model_name="gpt-3.5-turbo-1106")
chain = load_summarize_chain(llm, chain_type="stuff")
 
# Step 4: Run the summarization chain on the loaded document
summarized_content = chain.invoke(docs)
summarized = summarized_content["input_documents"][0].to_json()
 
# Step 5: Define the needed dependencies
dependencies = {
    "langchain": ">=1.0.0",
    "langchain_community": ">=1.0.0",
    "langchain_openai": ">=1.0.0"
}
 
result = chain.invoke(docs)
 
await ctx.send(
    sender,
    UAgentResponse(message=(result["output_text"]), type=UAgentResponseType.FINAL),
)
 
 
summaryAgent.include(summary_protocol, publish_manifest=True)
summaryAgent.run()

This Agent is simple again, hopefully you're seeing a pattern here. These Agents while independent of each other can compliment one another.

Aside from the different logic, the primary difference in web_summary_agent.py is that this agent does not rely on a separate function, and wraps the whole logic in the message handler.

Let's run this one too. Again, follow the steps for the previous Agent and go and register a Mailbox and update this agent with the new Mailbox. Each agent must have a unique Mailbox.

One more service 🎵

The following is the service for the Summary_agent:

This one is quite simple, but we are defining this as a subtask. A subtask tells AI engine that this function should only be called to support another agent, and not as a primary response. You can learn more about this in our services guide ↗️

We also need to go and update the sentiment service...

Oh my, that is quite the field description.

"This agent will return a sentiment score from the provided text. This MUST NOT be a URL or something that looks like a webpage address. If one of these is provided the content of that must be generated from another subtask. Text is the content to be analysed, this must be a structured body of prose."

We're talking to an LLM and the above should reinforce that. We're telling the AI Engine that this function returns sentiment analysis on text. However when this service is called, the user could also provide a URL. If a URL is provided then the AI engine must get that from somewhere else. Let's test this.

Let's go to DeltaV!

We're hoping here that we can ask for a sentiment of a webpage. You may need to adjust your question, or update field descriptions for this to work perfectly, but in latest this is working for us (24/4/2024).

What's next?

Thank you for sticking with us on this guide, this guide aims to show you the true potential of an Ai system that can execute code at the prompt of a user.

We invite you to dive deeper into the world of Agents, Agentverse and AI Engine by exploring the dedicated documentation and GitHub ↗️ (opens in a new tab) repository. Join our Discord ↗️ (opens in a new tab) and team up with other developers to participate in hackathons, collectively build projects, or simply have fun!