One of many attention-grabbing areas into these Agentic Ai methods is the distinction between constructing a single versus multi-agent workflow, or maybe the distinction between working with extra versatile vs managed methods.
This text will make it easier to perceive what agentic AI is, how one can construct easy workflows with LangGraph, and the variations in outcomes you’ll be able to obtain with the completely different architectures. I’ll show this by constructing a tech information agent with varied knowledge sources.
As for the use case, I’m a bit obsessive about getting automated information updates, primarily based on my preferences, with out me drowning in data overload on daily basis.
Working with summarizing and gathering analysis is a type of areas that agentic AI can actually shine.
So observe alongside whereas I preserve attempting to make AI do the grunt work for me, and we’ll see how single-agent compares to multi-agent setups.
I at all times preserve my work jargon-free, so when you’re new to agentic AI, this piece ought to make it easier to perceive what it’s and how one can work with it. If you happen to’re not new to it, you’ll be able to scroll previous a number of the sections.
Agentic AI (& LLMs)
Agentic AI is about programming with pure language. As an alternative of utilizing inflexible, express code, you’re instructing massive language fashions (LLMs) to route knowledge and carry out actions by means of plain language in automating duties.
Utilizing pure language in workflows isn’t new, we’ve used NLP for years to extract and course of knowledge. What’s new is the quantity of freedom we will now give language fashions, permitting them to deal with ambiguity and make choices dynamically.
However simply because LLMs can perceive nuanced language doesn’t imply they inherently validate info or keep knowledge integrity. I see them primarily as a communication layer that sits on high of structured methods and current knowledge sources.
I normally clarify it like this to non-technical individuals: they work a bit like we do. If we don’t have entry to scrub, structured knowledge, we begin making issues up. Identical with LLMs. They generate responses primarily based on patterns, not truth-checking.
So identical to us, they do their greatest with what they’ve acquired. If we would like higher output, we have to construct methods that give them dependable knowledge to work with. So, with Agentic methods we combine methods for them to work together with completely different knowledge sources, instruments and methods.
Now, simply because we can use these bigger fashions in additional locations, doesn’t imply we ought to. LLMs shine when decoding nuanced pure language, suppose customer support, analysis, or human-in-the-loop collaboration.
However for structured duties — like extracting numbers and sending them someplace — it’s good to use conventional approaches. LLMs aren’t inherently higher at math than a calculator. So, as a substitute of getting an LLM do calculations, you give an LLM entry to a calculator.
So everytime you can construct components of a workflow programmatically, that may nonetheless be the higher possibility.
However, LLMs are nice at adapting to messy real-world enter and decoding imprecise directions so combining the 2 may be an effective way to construct methods.
Agentic Frameworks
I do know lots of people leap straight to CrewAI or AutoGen right here, however I’d suggest trying out LangGraph, Agno, Mastra, and Smolagents. Based mostly on my analysis, these frameworks have acquired a number of the strongest suggestions to this point.
LangGraph is extra technical and may be advanced, but it surely’s the popular alternative for a lot of builders. Agno is simpler to get began with however much less technical. Mastra is a stable possibility for JavaScript builders, and Smolagents reveals lots of promise as a light-weight different.
On this case, I’ve gone with LangGraph — constructed on high of LangChain — not as a result of it’s my favourite, however as a result of it’s changing into a go-to framework that extra devs are adopting.
So, it’s value being acquainted with.
It has lots of abstractions although, the place it’s possible you’ll need to rebuild a few of it simply to have the ability to management and perceive it higher.
I cannot go into element on LangGraph right here, so I made a decision to construct a fast guide for those who have to get a evaluation.
As for this use case, you’ll have the ability to run the workflow with out coding something, however when you’re right here to study you might also need to perceive the way it works.
Selecting an LLM
Now, you may leap into this and marvel why I’m selecting sure LLMs as the bottom for the brokers.
You may’t simply decide any mannequin, particularly when working inside a framework. They have to be appropriate. Key issues to search for are device calling help and the flexibility to generate structured outputs.
I’d suggest checking HuggingFace’s Agent Leaderboard to see which fashions really carry out effectively in real-world agentic methods.
For this workflow, you have to be advantageous utilizing fashions from Anthropic, OpenAI, or Google. If you happen to’re contemplating one other one, simply make sure that it’s appropriate with LangChain.
Single vs. Multi-Agent Programs
If you happen to construct a system round one LLM and provides it a bunch of instruments you need it to make use of, you’re working with a single-agent workflow. It’s quick, and when you’re new to agentic AI, it would seem to be the mannequin ought to simply determine issues out by itself.
However the factor is these workflows are simply one other type of system design. Like all software program mission, it’s good to plan the method, outline the steps, construction the logic, and determine how every half ought to behave.
That is the place multi-agent workflows are available in.
Not all of them are hierarchical or linear although, some are collaborative. Collaborative workflows would then additionally fall into the extra versatile method that I discover tougher to work with, at the very least as it’s now with the capabilities that exist.
Nonetheless, collaborative workflows do additionally break aside completely different capabilities into their very own modules.
Single-agent and collaborative workflows are nice to start out with while you’re simply taking part in round, however they don’t at all times provide the precision wanted for precise duties.
For the workflow I’ll construct right here, I already know the way the APIs ought to be used — so it’s my job to information the system to make use of it the appropriate method.
We’ll undergo evaluating a single-agent setup with a hierarchical multi-agent system, the place a lead agent delegates duties throughout a small group so you’ll be able to see how they behave in observe.
Constructing a Single Agent Workflow
With a single thread — i.e., one agent — we give an LLM entry to a number of instruments. It’s as much as the agent to determine which device to make use of and when, primarily based on the person’s query.
The problem with a single agent is management.
Regardless of how detailed the system immediate is, the mannequin could not observe our requests (this may occur in additional managed environments too). If we give it too many instruments or choices, there’s an excellent probability it received’t use all of them and even use the appropriate ones.
For instance this, we’ll construct a tech information agent that has entry to a number of API endpoints with customized knowledge with a number of choices as parameters within the instruments. It’s as much as the agent to determine what number of to make use of and how one can setup the ultimate abstract.
Bear in mind, I construct these workflows utilizing LangGraph. I received’t go into LangGraph in depth right here, so if you wish to study the fundamentals to have the ability to tweak the code, go here.
You could find the single-agent workflow here. To run it, you’ll want LangGraph Studio and the newest model of Docker put in.
When you’re arrange, open the mission folder in your pc, add your GOOGLE_API_KEY in a .env file, and save. You will get a key from Google here.
Gemini Flash 2.0 has a beneficiant free tier, so working this shouldn’t value something (however it’s possible you’ll run into errors when you use it an excessive amount of).
If you wish to swap to a different LLM or instruments, you’ll be able to tweak the code immediately. However, once more, keep in mind the LLM must be appropriate.
After setup, launch LangGraph Studio and choose the right folder.
It will boot up our workflow so we will take a look at it.
If you happen to run into points booting this up, double-check that you simply’re utilizing the newest model of Docker.
As soon as it’s loaded, you’ll be able to take a look at the workflow by coming into a human message and hitting submit.
You may see me run the workflow beneath.
You may see the ultimate response beneath.
For this immediate it determined that it might verify weekly trending key phrases filtered by the class ‘corporations’ solely, after which it fetched the sources of these key phrases and summarized for us.
It had some points in giving us a unified abstract, the place it merely used the knowledge it acquired final and failed to make use of all the analysis.
In actuality we would like it to fetch each trending and high key phrases inside a number of classes (not simply corporations), verify sources, observe particular key phrases, and cause and summarize all of it properly earlier than returning a response.
We will in fact probe it and preserve asking it questions however as you’ll be able to think about if we want one thing extra advanced it might begin to make shortcuts within the workflow.
The important thing factor is, an agent system isn’t simply gonna suppose the way in which we anticipate, we’ve got to truly orchestrate it to do what we would like.
So a single agent is nice for one thing easy however as you’ll be able to think about it could not suppose or behave as we predict.
That is why going for a extra advanced system the place every agent is chargeable for one factor may be actually helpful.
Testing a Multi-Agent Workflow
Constructing multiagent workflows is much more troublesome than constructing a single agent with entry to some instruments. To do that, it’s good to rigorously take into consideration the structure beforehand and the way knowledge ought to stream between the brokers.
The multi-agent workflow I’ll arrange right here makes use of two completely different groups — a analysis group and an enhancing group — with a number of brokers below every.
Each agent has entry to a particular set of instruments.
We’re introducing some new instruments, like a analysis pad that acts as a shared house — one group writes their findings, the opposite reads from it. The final LLM will learn every little thing that has been researched and edited to make a abstract.
A substitute for utilizing a analysis pad is to retailer knowledge in a scratchpad in state, isolating short-term reminiscence for every group or agent. However that additionally means considering rigorously about what every agent’s reminiscence ought to embody.
I additionally determined to construct out the instruments a bit extra to offer richer knowledge upfront, so the brokers don’t should fetch sources for every key phrase individually. Right here I’m utilizing regular programmatic logic as a result of I can.
A key factor to recollect: if you should utilize regular programming logic, do it.
Since we’re utilizing a number of brokers, you’ll be able to decrease prices through the use of cheaper fashions for many brokers and reserving the costlier ones for the essential stuff.
Right here, I’m utilizing Gemini Flash 2.0 for all brokers besides the summarizer, which runs on OpenAI’s GPT-4o. If you would like higher-quality summaries, you should utilize an much more superior LLM with a bigger context window.
The workflow is about up for you here. Earlier than loading it, make sure that so as to add each your OpenAI and Google API keys in a .env file.
On this workflow, the routes (edges) are setup dynamically as a substitute of manually like we did with the one agent. It’ll look extra advanced when you peek into the code.
When you boot up the workflow in LangGraph Studio — similar course of as earlier than — you’ll see the graph with all these nodes prepared.
LangGraph Studio lets us visualize how the system delegates work between brokers after we run it—identical to we noticed within the easier workflow above.
Since I perceive the instruments every agent is utilizing, I can immediate the system in the appropriate method. However common customers received’t know the way to do that correctly. So when you’re constructing one thing related, I’d recommend introducing an agent that transforms the person’s question into one thing the opposite brokers can really work with.
We will check it out by setting a message.
“I’m an investor and I’m enthusiastic about getting an replace for what has occurred inside the week in tech, and what persons are speaking about (this implies classes like corporations, individuals, web sites and topics are attention-grabbing). Please additionally observe these particular key phrases: AI, Google, Microsoft, and Giant Language Fashions”
Then selecting “supervisor” because the Subsequent parameter (we’d usually do that programmatically).
This workflow will take a number of minutes to run, in contrast to the single-agent workflow we ran earlier which completed in below a minute.
So be affected person whereas the instruments are working.
Normally, these methods take time to collect and course of data and that’s simply one thing we have to get used to.
The ultimate abstract will look one thing like this:
You may learn the entire thing here as a substitute if you wish to test it out.
The information will clearly differ relying on while you run the workflow. I ran it the twenty eighth of March so the instance report can be for this date.
It ought to save the abstract to a textual content doc, however when you’re working this inside a container, you doubtless received’t have the ability to entry that file simply. It’s higher to ship the output some other place — like Google Docs or by way of e-mail.
As for the outcomes, I’ll allow you to determine for your self the distinction between utilizing a extra advanced system versus a easy one, and the way it provides us extra management over the method.
Ending Notes
I’m working with an excellent knowledge supply right here. With out that, you’d want so as to add much more error dealing with, which might gradual every little thing down much more.
Clear and structured knowledge is vital. With out it, the LLM received’t carry out at its greatest.
Even with stable knowledge, it’s not excellent. You continue to have to work on the brokers to ensure they do what they’re presupposed to.
You’ve in all probability already seen the system works — but it surely’s not fairly there but.
There are nonetheless a number of issues that want enchancment: parsing the person’s question right into a extra structured format, including guardrails so brokers at all times use their instruments, summarizing extra successfully to maintain the analysis doc concise, enhancing error dealing with, and introducing long-term reminiscence to higher perceive what the person really wants.
State (short-term reminiscence) is very essential if you wish to optimize for efficiency and value.
Proper now, we’re simply pushing each message into state and giving all brokers entry to it, which isn’t excellent. We actually need to separate state between the groups. On this case, it’s one thing I haven’t carried out, however you’ll be able to attempt it by introducing a scratchpad within the state schema to isolate what every group is aware of.
Regardless, I hope it was a enjoyable expertise to grasp the outcomes we will get by constructing completely different Agentic Workflows.
If you wish to see extra of what I’m engaged on, you’ll be able to observe me here but additionally on Medium, GitHub, or LinkedIn (although I’m hoping to maneuver over to X quickly). I even have a Substack, the place I hope to publishing shorter items in.
❤️
