The tradeoffs between constructing bespoke code-based brokers and the key agent frameworks.
Because of John Gilhuly for his contributions to this piece.
Brokers are having a second. With a number of new frameworks and recent funding within the area, trendy AI brokers are overcoming shaky origins to quickly supplant RAG as an implementation precedence. So will 2024 lastly be the yr that autonomous AI programs that may take over writing our emails, reserving flights, speaking to our information, or seemingly some other activity?
Possibly, however a lot work stays to get to that time. Any developer constructing an agent should not solely select foundations — which mannequin, use case, and structure to make use of — but additionally which framework to leverage. Do you go together with the long-standing LangGraph, or the newer entrant LlamaIndex Workflows? Or do you go the standard route and code the entire thing your self?
This publish goals to make that selection a bit simpler. Over the previous few weeks, I constructed the identical agent in main frameworks to look at a number of the strengths and weaknesses of every at a technical stage. All the code for every agent is offered in this repo.
Background on the Agent Used for Testing
The agent used for testing contains perform calling, a number of instruments or expertise, connections to outdoors assets, and shared state or reminiscence.
The agent has the next capabilities:
- Answering questions from a data base
- Speaking to information: answering questions on telemetry information of an LLM software
- Analyzing information: analyzing higher-level tendencies and patterns in retrieved telemetry information
To be able to accomplish these, the agent has three beginning expertise: RAG with product documentation, SQL technology on a hint database, and information evaluation. A easy gradio-powered interface is used for the agent UI, with the agent itself structured as a chatbot.
The primary choice you will have when creating an agent is to skip the frameworks fully and construct the agent absolutely your self. When embarking on this undertaking, this was the strategy I began with.
Pure Code Structure
The code-based agent under is made up of an OpenAI-powered router that makes use of perform calling to pick the correct talent to make use of. After that talent completes, it returns again to the router to both name one other talent or reply to the consumer.
The agent retains an ongoing listing of messages and responses that’s handed absolutely into the router on every name to protect context by way of cycles.
def router(messages):
if not any(
isinstance(message, dict) and message.get("position") == "system" for message in messages
):
system_prompt = {"position": "system", "content material": SYSTEM_PROMPT}
messages.append(system_prompt)response = consumer.chat.completions.create(
mannequin="gpt-4o",
messages=messages,
instruments=skill_map.get_combined_function_description_for_openai(),
)
messages.append(response.selections[0].message)
tool_calls = response.selections[0].message.tool_calls
if tool_calls:
handle_tool_calls(tool_calls, messages)
return router(messages)
else:
return response.selections[0].message.content material
The talents themselves are outlined in their very own courses (e.g. GenerateSQLQuery) which can be collectively held in a SkillMap. The router itself solely interacts with the SkillMap, which it makes use of to load talent names, descriptions, and callable capabilities. This strategy implies that including a brand new talent to the agent is so simple as writing that talent as its personal class, then including it to the listing of expertise within the SkillMap. The concept right here is to make it straightforward so as to add new expertise with out disturbing the router code.
class SkillMap:
def __init__(self):
expertise = [AnalyzeData(), GenerateSQLQuery()]self.skill_map = {}
for talent in expertise:
self.skill_map[skill.get_function_name()] = (
talent.get_function_dict(),
talent.get_function_callable(),
)
def get_function_callable_by_name(self, skill_name) -> Callable:
return self.skill_map[skill_name][1]
def get_combined_function_description_for_openai(self):
combined_dict = []
for _, (function_dict, _) in self.skill_map.objects():
combined_dict.append(function_dict)
return combined_dict
def get_function_list(self):
return listing(self.skill_map.keys())
def get_list_of_function_callables(self):
return [skill[1] for talent in self.skill_map.values()]
def get_function_description_by_name(self, skill_name):
return str(self.skill_map[skill_name][0]["function"])
Total, this strategy is pretty simple to implement however comes with a couple of challenges.
Challenges with Pure Code Brokers
The primary issue lies in structuring the router system immediate. Typically, the router within the instance above insisted on producing SQL itself as an alternative of delegating that to the correct talent. For those who’ve ever tried to get an LLM not to do one thing, you understand how irritating that have could be; discovering a working immediate took many rounds of debugging. Accounting for the completely different output codecs from every step was additionally tough. Since I opted to not use structured outputs, I needed to be prepared for a number of completely different codecs from every of the LLM calls in my router and expertise.
Advantages of a Pure Code Agent
A code-based strategy gives baseline and place to begin, providing an effective way to learn the way brokers work with out counting on canned agent tutorials from prevailing frameworks. Though convincing the LLM to behave could be difficult, the code construction itself is straightforward sufficient to make use of and would possibly make sense for sure use circumstances (extra within the evaluation part under).
LangGraph is likely one of the longest-standing agent frameworks, first releasing in January 2024. The framework is constructed to deal with the acyclic nature of current pipelines and chains by adopting a Pregel graph construction as an alternative. LangGraph makes it simpler to outline loops in your agent by including the ideas of nodes, edges, and conditional edges to traverse a graph. LangGraph is constructed on high of LangChain, and makes use of the objects and kinds from that framework.
LangGraph Structure
The LangGraph agent seems just like the code-based agent on paper, however the code behind it’s drastically completely different. LangGraph nonetheless makes use of a “router” technically, in that it calls OpenAI with capabilities and makes use of the response to proceed to a brand new step. Nonetheless the way in which this system strikes between expertise is managed fully in a different way.
instruments = [generate_and_run_sql_query, data_analyzer]
mannequin = ChatOpenAI(mannequin="gpt-4o", temperature=0).bind_tools(instruments)def create_agent_graph():
workflow = StateGraph(MessagesState)
tool_node = ToolNode(instruments)
workflow.add_node("agent", call_model)
workflow.add_node("instruments", tool_node)
workflow.add_edge(START, "agent")
workflow.add_conditional_edges(
"agent",
should_continue,
)
workflow.add_edge("instruments", "agent")
checkpointer = MemorySaver()
app = workflow.compile(checkpointer=checkpointer)
return app
The graph outlined right here has a node for the preliminary OpenAI name, referred to as “agent” above, and one for the device dealing with step, referred to as “instruments.” LangGraph has a built-in object referred to as ToolNode that takes a listing of callable instruments and triggers them primarily based on a ChatMessage response, earlier than returning to the “agent” node once more.
def should_continue(state: MessagesState):
messages = state["messages"]
last_message = messages[-1]
if last_message.tool_calls:
return "instruments"
return ENDdef call_model(state: MessagesState):
messages = state["messages"]
response = mannequin.invoke(messages)
return {"messages": [response]}
After every name of the “agent” node (put one other means: the router within the code-based agent), the should_continue edge decides whether or not to return the response to the consumer or move on to the ToolNode to deal with device calls.
All through every node, the “state” shops the listing of messages and responses from OpenAI, just like the code-based agent’s strategy.
Challenges with LangGraph
Many of the difficulties with LangGraph within the instance stem from the necessity to use Langchain objects for issues to circulate properly.
Problem #1: Perform Name Validation
To be able to use the ToolNode object, I needed to refactor most of my current Talent code. The ToolNode takes a listing of callable capabilities, which initially made me suppose I might use my current capabilities, nonetheless issues broke down because of my perform parameters.
The talents had been outlined as courses with a callable member perform, which means they’d “self” as their first parameter. GPT-4o was good sufficient to not embrace the “self” parameter within the generated perform name, nonetheless LangGraph learn this as a validation error because of a lacking parameter.
This took hours to determine, as a result of the error message as an alternative marked the third parameter within the perform (“args” on the information evaluation talent) because the lacking parameter:
pydantic.v1.error_wrappers.ValidationError: 1 validation error for data_analysis_toolSchema
args discipline required (kind=value_error.lacking)
It’s price mentioning that the error message originated from Pydantic, not from LangGraph.
I finally bit the bullet and redefined my expertise as primary strategies with Langchain’s @device decorator, and was in a position to get issues working.
@device
def generate_and_run_sql_query(question: str):
"""Generates and runs an SQL question primarily based on the immediate.Args:
question (str): A string containing the unique consumer immediate.
Returns:
str: The results of the SQL question.
"""
Problem #2: Debugging
As talked about, debugging in a framework is tough. This primarily comes right down to complicated error messages and abstracted ideas that make it more durable to view variables.
The abstracted ideas primarily present up when making an attempt to debug the messages being despatched across the agent. LangGraph shops these messages in state[“messages”]. Some nodes throughout the graph pull from these messages mechanically, which may make it obscure the worth of messages when they’re accessed by the node.
LangGraph Advantages
One of many fundamental advantages of LangGraph is that it’s straightforward to work with. The graph construction code is clear and accessible. Particularly when you’ve got advanced node logic, having a single view of the graph makes it simpler to know how the agent is linked collectively. LangGraph additionally makes it simple to transform an current software inbuilt LangChain.
Takeaway
For those who use the whole lot within the framework, LangGraph works cleanly; in the event you step outdoors of it, put together for some debugging complications.
Workflows is a more recent entrant into the agent framework area, premiering earlier this summer season. Like LangGraph, it goals to make looping brokers simpler to construct. Workflows additionally has a selected deal with working asynchronously.
Some components of Workflows appear to be in direct response to LangGraph, particularly its use of occasions as an alternative of edges and conditional edges. Workflows use steps (analogous to nodes in LangGraph) to deal with logic, and emitted and acquired occasions to maneuver between steps.
The construction above seems just like the LangGraph construction, save for one addition. I added a setup step to the Workflow to arrange the agent context, extra on this under. Regardless of the same construction, there may be very completely different code powering it.
Workflows Structure
The code under defines the Workflow construction. Just like LangGraph, that is the place I ready the state and connected the abilities to the LLM object.
class AgentFlow(Workflow):
def __init__(self, llm, timeout=300):
tremendous().__init__(timeout=timeout)
self.llm = llm
self.reminiscence = ChatMemoryBuffer(token_limit=1000).from_defaults(llm=llm)
self.instruments = []
for func in skill_map.get_function_list():
self.instruments.append(
FunctionTool(
skill_map.get_function_callable_by_name(func),
metadata=ToolMetadata(
identify=func, description=skill_map.get_function_description_by_name(func)
),
)
)@step
async def prepare_agent(self, ev: StartEvent) -> RouterInputEvent:
user_input = ev.enter
user_msg = ChatMessage(position="consumer", content material=user_input)
self.reminiscence.put(user_msg)
chat_history = self.reminiscence.get()
return RouterInputEvent(enter=chat_history)
That is additionally the place I outline an additional step, “prepare_agent”. This step creates a ChatMessage from the consumer enter and provides it to the workflow reminiscence. Splitting this out as a separate step implies that we do return to it because the agent loops by way of steps, which avoids repeatedly including the consumer message to the reminiscence.
Within the LangGraph case, I completed the identical factor with a run_agent technique that lived outdoors the graph. This alteration is usually stylistic, nonetheless it’s cleaner in my view to deal with this logic with the Workflow and graph as we’ve achieved right here.
With the Workflow arrange, I then outlined the routing code:
@step
async def router(self, ev: RouterInputEvent) -> ToolCallEvent | StopEvent:
messages = ev.enterif not any(
isinstance(message, dict) and message.get("position") == "system" for message in messages
):
system_prompt = ChatMessage(position="system", content material=SYSTEM_PROMPT)
messages.insert(0, system_prompt)
with using_prompt_template(template=SYSTEM_PROMPT, model="v0.1"):
response = await self.llm.achat_with_tools(
mannequin="gpt-4o",
messages=messages,
instruments=self.instruments,
)
self.reminiscence.put(response.message)
tool_calls = self.llm.get_tool_calls_from_response(response, error_on_no_tool_call=False)
if tool_calls:
return ToolCallEvent(tool_calls=tool_calls)
else:
return StopEvent(outcome=response.message.content material)
And the device name dealing with code:
@step
async def tool_call_handler(self, ev: ToolCallEvent) -> RouterInputEvent:
tool_calls = ev.tool_callsfor tool_call in tool_calls:
function_name = tool_call.tool_name
arguments = tool_call.tool_kwargs
if "enter" in arguments:
arguments["prompt"] = arguments.pop("enter")
strive:
function_callable = skill_map.get_function_callable_by_name(function_name)
besides KeyError:
function_result = "Error: Unknown perform name"
function_result = function_callable(arguments)
message = ChatMessage(
position="device",
content material=function_result,
additional_kwargs={"tool_call_id": tool_call.tool_id},
)
self.reminiscence.put(message)
return RouterInputEvent(enter=self.reminiscence.get())
Each of those look extra just like the code-based agent than the LangGraph agent. That is primarily as a result of Workflows retains the conditional routing logic within the steps versus in conditional edges — traces 18–24 had been a conditional edge in LangGraph, whereas now they’re simply a part of the routing step — and the truth that LangGraph has a ToolNode object that does nearly the whole lot within the tool_call_handler technique mechanically.
Shifting previous the routing step, one factor I used to be very joyful to see is that I might use my SkillMap and current expertise from my code-based agent with Workflows. These required no modifications to work with Workflows, which made my life a lot simpler.
Challenges with Workflows
Problem #1: Sync vs Async
Whereas asynchronous execution is preferable for a dwell agent, debugging a synchronous agent is far simpler. Workflows is designed to work asynchronously, and making an attempt to drive synchronous execution was very tough.
I initially thought I might simply be capable of take away the “async” technique designations and swap from “achat_with_tools” to “chat_with_tools”. Nonetheless, because the underlying strategies throughout the Workflow class had been additionally marked as asynchronous, it was essential to redefine these so as to run synchronously. I ended up sticking to an asynchronous strategy, however this didn’t make debugging tougher.
Problem #2: Pydantic Validation Errors
In a repeat of the woes with LangGraph, related issues emerged round complicated Pydantic validation errors on expertise. Fortuitously, these had been simpler to deal with this time since Workflows was in a position to deal with member capabilities simply fantastic. I finally simply ended up having to be extra prescriptive in creating LlamaIndex FunctionTool objects for my expertise:
for func in skill_map.get_function_list():
self.instruments.append(FunctionTool(
skill_map.get_function_callable_by_name(func),
metadata=ToolMetadata(identify=func, description=skill_map.get_function_description_by_name(func))))
Excerpt from AgentFlow.__init__ that builds FunctionTools
Advantages of Workflows
I had a a lot simpler time constructing the Workflows agent than I did the LangGraph agent, primarily as a result of Workflows nonetheless required me to write down routing logic and power dealing with code myself as an alternative of offering built-in capabilities. This additionally meant that my Workflow agent regarded extraordinarily just like my code-based agent.
The largest distinction got here in using occasions. I used two customized occasions to maneuver between steps in my agent:
class ToolCallEvent(Occasion):
tool_calls: listing[ToolSelection]class RouterInputEvent(Occasion):
enter: listing[ChatMessage]
The emitter-receiver, event-based structure took the place of straight calling a number of the strategies in my agent, just like the device name handler.
When you’ve got extra advanced programs with a number of steps which can be triggering asynchronously and would possibly emit a number of occasions, this structure turns into very useful to handle that cleanly.
Different advantages of Workflows embrace the truth that it is rather light-weight and doesn’t drive a lot construction on you (other than using sure LlamaIndex objects) and that its event-based structure gives a useful different to direct perform calling — particularly for advanced, asynchronous purposes.
Trying throughout the three approaches, each has its advantages.
The no framework strategy is the best to implement. As a result of any abstractions are outlined by the developer (i.e. SkillMap object within the above instance), preserving varied sorts and objects straight is simple. The readability and accessibility of the code fully comes right down to the person developer nonetheless, and it’s straightforward to see how more and more advanced brokers might get messy with out some enforced construction.
LangGraph gives fairly a little bit of construction, which makes the agent very clearly outlined. If a broader group is collaborating on an agent, this construction would offer a useful means of implementing an structure. LangGraph additionally would possibly present place to begin with brokers for these not as aware of the construction. There’s a tradeoff, nonetheless — since LangGraph does fairly a bit for you, it may possibly result in complications in the event you don’t absolutely purchase into the framework; the code could also be very clear, however it’s possible you’ll pay for it with extra debugging.
Workflows falls someplace within the center. The event-based structure could be extraordinarily useful for some tasks, and the truth that much less is required when it comes to utilizing of LlamaIndex sorts gives larger flexibility for these not be absolutely utilizing the framework throughout their software.
In the end, the core query could come right down to “are you already utilizing LlamaIndex or LangChain to orchestrate your software?” LangGraph and Workflows are each so entwined with their respective underlying frameworks that the extra advantages of every agent-specific framework may not trigger you to modify on advantage alone.
The pure code strategy will probably at all times be a lovely choice. When you’ve got the rigor to doc and implement any abstractions created, then guaranteeing nothing in an exterior framework slows you down is simple.
In fact, “it relies upon” isn’t a satisfying reply. These three questions ought to enable you to determine which framework to make use of in your subsequent agent undertaking.
Are you already utilizing LlamaIndex or LangChain for important items of your undertaking?
If sure, discover that choice first.
Are you aware of widespread agent buildings, or would you like one thing telling you the way it is best to construction your agent?
For those who fall into the latter group, strive Workflows. For those who actually fall into the latter group, strive LangGraph.
Has your agent been constructed earlier than?
One of many framework advantages is that there are numerous tutorials and examples constructed with every. There are far fewer examples of pure code brokers to construct from.
Selecting an agent framework is only one selection amongst many that can impression outcomes in manufacturing for generative AI programs. As at all times, it pays to have strong guardrails and LLM tracing in place — and to be agile as new agent frameworks, analysis, and fashions upend established strategies.