Why CrewAI’s Manager-Worker Architecture Fails — and How to Fix It

is without doubt one of the most promising purposes of LLMs, and CrewAI has shortly change into a well-liked framework for constructing agent groups. However considered one of its most necessary options—the hierarchical manager-worker course of—merely doesn’t perform as documented. In actual workflows, the supervisor doesn’t successfully coordinate brokers; as an alternative, CrewAI executes duties sequentially, resulting in incorrect reasoning, pointless instrument calls, and intensely excessive latency. This subject has been highlighted in a number of on-line boards with no clear decision.

On this article, I exhibit why CrewAI’s hierarchical course of fails, present the proof from precise Langfuse traces, and supply a reproducible pathway to make the manager-worker sample work reliably utilizing customized prompting.

Multi-agent Orchestration

Earlier than we get into the main points, allow us to perceive what orchestration means in an agentic context. In easy phrases, orchestration is managing and coordinating a number of inter-dependent duties in a workflow. However have’nt workflow administration instruments (eg; RPA) been out there endlessly to do exactly that? So what modified with LLMs?

The reply is the flexibility of LLMs to grasp that means and intent from pure language directions, simply as folks in a crew would. Whereas earlier workflow instruments have been rule-based and inflexible, with LLMs functioning as brokers, the expectation is that they are going to be capable of perceive the intent of the person’s question, use reasoning to create a multi-step plan, infer the instruments for use, derive their inputs within the right codecs, and synthesize all of the totally different intermediate ends in a exact response to the person’s question. And the orchestration frameworks are supposed to information the LLM with acceptable prompts for planning, tool-calling, producing response and so forth.

Among the many orchestration frameworks, CrewAI, with its pure language primarily based definition of duties, brokers and crews relies upon essentially the most on the LLM’s skill to grasp language and handle workflows. Whereas not as deterministic as LangGraph (since LLM outputs can’t be totally deterministic), it abstracts away many of the complexity of routing, error dealing with and so forth into easy, user-friendly constructs with parameters, which the person can tune for acceptable conduct. So it’s a good framework for creating prototypes by product groups and even non-developers.

Besides that the manager-worker sample doesn’t work as meant…

As an example, let’s take a use-case to work with. And likewise consider the response primarily based on the next standards:

1. High quality of orchestration
2. High quality of ultimate response
3. Explainability
4. Latency and utilization price

Use Case

Take the case the place a crew of buyer help brokers resolve technical or billing tickets. When a ticket comes, a triage agent categorizes the ticket, then assigns to the technical or billing help specialist for decision. There’s a Buyer Assist Supervisor who coordinates the crew, delegates duties and validates high quality of response.

Collectively they are going to be fixing queries corresponding to:

1. Why is my laptop computer overheating?
2. Why was I charged twice final month?
3. My laptop computer is overheating and likewise, I used to be charged twice final month?
4. My bill quantity is inaccurate after system glitch?

The primary question is solely technical, so solely the technical help agent needs to be invoked by the supervisor, the second is Billing solely and the third and fourth ones require solutions from each technical and billing brokers.

Let’s construct this crew of CrewAI brokers and see how nicely it really works.

Hierarchical Course of

In response to CrewAI documentation ,“adopting a hierarchical method permits for a transparent hierarchy in job administration, the place a ‘supervisor’ agent coordinates the workflow, delegates duties, and validates outcomes for streamlined and efficient execution. “ Additionally, the supervisor agent may be created in two methods, mechanically by CrewAI or explicitly set by the person. Within the latter case, you’ve got extra management over directions to the supervisor agent. We are going to strive each methods for our use case.

CrewAI Code

from crewai import Agent, Crew, Course of, Process, LLM
from dotenv import load_dotenv
import os
from observe import * # Langfuse hint

load_dotenv()
verbose = False
max_iter = 4

API_VERSION = os.getenv(API_VERSION')
# Create your LLM
llm_a = LLM(
    mannequin="gpt-4o",
    api_version=API_VERSION,
    temperature = 0.2,
    max_tokens = 8000,
)

# Outline the supervisor agent
supervisor = Agent(
    position="Buyer Assist Supervisor",
    purpose="Oversee the help crew to make sure well timed and efficient decision of buyer inquiries. Use the instrument to categorize the person question first, then resolve the following steps.Syntesize responses from totally different brokers if wanted to supply a complete reply to the client.",
    backstory=( """
        You don't attempt to discover a solution to the person ticket {ticket} your self. 
        You delegate duties to coworkers primarily based on the next logic:
        Be aware the class of the ticket first through the use of the triage agent.
        If the ticket is categorized as 'Each', at all times assign it first to the Technical Assist Specialist, then to the Billing Assist Specialist, then print the ultimate mixed response. Make sure that the ultimate response solutions each technical and billing points raised within the ticket primarily based on the responses from each Technical and Billing Assist Specialists.
        ELSE
        If the ticket is categorized as 'Technical', assign it to the Technical Assist Specialist, else skip this step.
        Earlier than continuing additional, analyse the ticket class. Whether it is 'Technical', print the ultimate response. Terminate additional actions.
        ELSE
        If the ticket is categorized as 'Billing', assign it to the Billing Assist Specialist.
        Lastly, compile and current the ultimate response to the client primarily based on the outputs from the assigned brokers.
        """
    ),
    llm = llm_a,
    allow_delegation=True,
    verbose=verbose,
)

# Outline the triage agent
triage_agent = Agent(
    position="Question Triage Specialist",
    purpose="Categorize the person question into technical or billing associated points. If a question requires each elements, reply with 'Each'.",
    backstory=(
        "You're a seasoned knowledgeable in analysing intent of person question. You reply exactly with one phrase: 'Technical', 'Billing' or 'Each'."
    ),
    llm = llm_a,
    allow_delegation=False,
    verbose=verbose,
)

# Outline the technical help agent
technical_support_agent = Agent(
    position="Technical Assist Specialist",
    purpose="Resolve technical points reported by clients promptly and successfully",
    backstory=(
        "You're a extremely expert technical help specialist with a powerful background in troubleshooting software program and {hardware} points. "
        "Your major accountability is to help clients in resolving technical issues, making certain their satisfaction and the sleek operation of their merchandise."
    ),
    llm = llm_a,
    allow_delegation=False,
    verbose=verbose,
)

# Outline the billing help agent
billing_support_agent = Agent(
    position="Billing Assist Specialist",
    purpose="Tackle buyer inquiries associated to billing, funds, and account administration",
    backstory=(
        "You might be an skilled billing help specialist with experience in dealing with buyer billing inquiries. "
        "Your major goal is to supply clear and correct info concerning billing processes, resolve fee points, and help with account administration to make sure buyer satisfaction."
    ),
    llm = llm_a,
    allow_delegation=False,
    verbose=verbose,
)

# Outline duties
categorize_tickets = Process(
    description="Categorize the incoming buyer help ticket: '{ticket} primarily based on its content material to find out whether it is technical or billing-related. If a question requires each elements, reply with 'Each'.",
    expected_output="A categorized ticket labeled as 'Technical' or 'Billing' or 'Each'. Don't be verbose, simply reply with one phrase.",
    agent=triage_agent,
)

resolve_technical_issues = Process(
    description="Resolve technical points described within the ticket: '{ticket}'",
    expected_output="Detailed options supplied to every technical subject.",
    agent=technical_support_agent,
)

resolve_billing_issues = Process(
    description="Resolve billing points described within the ticket: '{ticket}'",
    expected_output="Complete responses to every billing-related inquiry.",
    agent=billing_support_agent,
)

# Instantiate your crew with a customized supervisor and hierarchical course of
crew_q = Crew(
    brokers=[triage_agent, technical_support_agent, billing_support_agent],
    duties=[categorize_tickets, resolve_technical_issues, resolve_billing_issues],
    # manager_llm = llm_a, # Uncomment for auto-created supervisor
    manager_agent=supervisor, # Remark for auto-created supervisor
    course of=Course of.hierarchical,
    verbose=verbose,
)

As is obvious, this system displays the crew of human brokers. Not solely is there a manger, triage agent, technical and billing help agent, however the CrewAI objects corresponding to Agent, Process and Crew are self-evident of their that means and simple to visualise. One other remark is that there’s little or no python code and many of the reasoning, planning and conduct is pure language primarily based which relies upon upon the flexibility of the LLM to derive that means and intent from language, then motive and plan for the purpose.

A CrewAI code subsequently, scores excessive on ease of growth. It’s a low-code means of making a movement shortly with many of the heavy-lifting of the workflow being finished by the orchestration framework quite than the developer.

How nicely does it work?

As we’re testing the hierarchical course of, the method parameter is about to Course of.hierarchical within the Crew definition. We will strive totally different options of CrewAI as follows and measure efficiency:

1. Supervisor agent auto-created by CrewAI
2. Utilizing our customized supervisor agent

1. Auto-created supervisor agent

Enter question: Why is my laptop computer overheating?

Right here is the Langfuse hint:

The important thing observations are as follows:

1. First the output is “Primarily based on the supplied context, it appears there’s a misalignment between the character of the difficulty (laptop computer overheating) and its categorization as a billing concern. To make clear the connection, it could be necessary to find out if the client is requesting a refund for the laptop computer because of the overheating subject, disputing a cost associated to the acquisition or restore of the laptop computer, or in search of compensation for restore prices incurred because of the overheating…” For a question that was clearly a technical subject, it is a poor response.
2. Why does it occur? The left panel reveals that the execution first went to triage specialist, then to technical help after which unusually, to billing help specialist as nicely. The next graphic depicts this as nicely:

Wanting intently, we discover that the triage specialist appropriately recognized the ticket as “Technical” and the technical help agent gave an important reply as follows:

However then, as an alternative of stopping and replying with the above because the response, the Crew Supervisor went to the Billing help specialist and tried to discover a non-existent billing subject within the purely technical person question.

This resulted within the Billing agent’s response overwriting the Technical agent’s response, with the Crew Supervisor doing a sub-optimal job of validating the standard of the ultimate response towards the person’s question.

Why did it occur?

As a result of within the Crew job definition, I specified the duties as categorize_tickets, resolve_technical_issues, resolve_billing_issues and though the method is meant to be hierarchical, the Crew Supervisor doesn’t carry out any orchestration, as an alternative merely executing all of the duties sequentially.

crew_q = Crew(
    brokers=[triage_agent, technical_support_agent, billing_support_agent],
    duties=[categorize_tickets, resolve_technical_issues, resolve_billing_issues],
    manager_llm = llm_a,
    course of=Course of.hierarchical,
    verbose=verbose,
)

When you now ask a billing-related question, it can seem to offer an accurate reply just because the resolve_billing_issues is the final job within the sequence.

What a couple of question that requires each technical and billing help, corresponding to “My laptop computer is overheating and likewise I used to be charged twice final month?” On this case additionally, the triage agent appropriately categorizes the ticket sort as “Each”, and the technical and billing brokers give right solutions to their particular person queries, however the supervisor is unable to mix all of the responses right into a coherent reply to person’s question. As a substitute, the ultimate response solely considers the billing response since that’s the final job to be referred to as in sequence.

Latency and Utilization: As may be seen from the above picture, the Crew execution took nearly 38 secs and spent 15759 tokens. The ultimate output is just about 200 tokens. The remainder of the tokens have been spent in all of the considering, agent calling, producing intermediate responses and so forth – all to generate an unsatisfactory response on the finish. The efficiency may be categorised as “Poor”.

Analysis of this method

• High quality of orchestration: Poor
• High quality of ultimate output: Poor
• Explainability: Poor
• Latency and Utilization: Poor

However maybe, the above end result is because of the truth that we relied on CrewAI’s built-in supervisor, which didn’t have our customized directions. Due to this fact, in our subsequent method we exchange the CrewAI automated supervisor with our customized Supervisor agent, which has detailed directions on what to do in case of Technical, Billing or Each tickets.

2. Utilizing Customized Supervisor Agent

Our Buyer Assist Supervisor is outlined with the next very particular directions. Be aware that this requires some experimentation to get it working, and a generic supervisor immediate corresponding to that talked about within the CrewAI documentation will give the identical faulty outcomes because the built-in supervisor agent above.

    position="Buyer Assist Supervisor",
    purpose="Oversee the help crew to make sure well timed and efficient decision of buyer inquiries. Use the instrument to categorize the person question first, then resolve the following steps.Syntesize responses from totally different brokers if wanted to supply a complete reply to the client.",
    backstory=( """
        You don't attempt to discover a solution to the person ticket {ticket} your self. 
        You delegate duties to coworkers primarily based on the next logic:
        Be aware the class of the ticket first through the use of the triage agent.
        If the ticket is categorized as 'Each', at all times assign it first to the Technical Assist Specialist, then to the Billing Assist Specialist, then print the ultimate mixed response. Make sure that the ultimate response solutions each technical and billing points raised within the ticket primarily based on the responses from each Technical and Billing Assist Specialists.
        ELSE
        If the ticket is categorized as 'Technical', assign it to the Technical Assist Specialist, else skip this step.
        Earlier than continuing additional, analyse the ticket class. Whether it is 'Technical', print the ultimate response. Terminate additional actions.
        ELSE
        If the ticket is categorized as 'Billing', assign it to the Billing Assist Specialist.
        Lastly, compile and current the ultimate response to the client primarily based on the outputs from the assigned brokers.
        """

And within the Crew definition, we use the customized supervisor as an alternative of the built-in one:

crew_q = Crew(
    brokers=[triage_agent, technical_support_agent, billing_support_agent],
    duties=[categorize_tickets, resolve_technical_issues, resolve_billing_issues],
    # manager_llm = llm_a,
    manager_agent=supervisor,
    course of=Course of.hierarchical,
    verbose=verbose,
)

Let’s repeat the take a look at circumstances

Enter question: Why is my laptop computer overheating?

The hint is the next:

Crucial remark is that now for this technical question, the movement didn’t go to the Billing help specialist agent. The supervisor appropriately adopted directions, categorized the question as technical and stopped execution as soon as the Technical Assist Specialist had generated its response. From the response preview displayed, it’s evident that it’s a good response for the person question. Additionally, the latency is 24 secs and token utilization is 10k.

Enter question: Why was I charged twice final month?

The hint is as follows:

As may be seen, the supervisor appropriately skipped executing the Technical Assist Specialist, though that was earlier than the Billing agent within the Crew definition. As a substitute the response generated is of fine high quality from the Billing Assist Specialist solely. Latency is 16 secs and token utilization 7,700 solely

Enter question: My laptop computer is overheating and likewise, I used to be charged twice final month?

The hint reveals the Supervisor executed each Technical and Billing help brokers and supplied a mixed response.

Pricey Buyer,

Thanks for reaching out to us concerning the problems you're experiencing. We sincerely apologize for any inconvenience prompted. Under are the detailed options to handle your issues:

**1. Laptop computer Overheating Problem:**
   - **Test for Correct Air flow**: Guarantee your laptop computer is positioned on a tough, flat floor to permit correct airflow. Keep away from utilizing it on gentle surfaces like beds or couches that may block the vents. Think about using a laptop computer cooling pad or stand with built-in followers to enhance airflow.
   - **Clear the Laptop computer's Vents and Followers**: Mud and particles can accumulate within the vents and followers, proscribing airflow. Energy off the laptop computer, unplug it, and use a can of compressed air to softly blow out mud from the vents. If you're comfy, you'll be able to clear the interior followers and parts extra completely, or take the laptop computer to knowledgeable technician for inside cleansing.
   - **Monitor Working Functions and Processes**: Open the Process Supervisor (Home windows: Ctrl + Shift + Esc, macOS: Exercise Monitor) and examine for processes consuming excessive CPU or GPU utilization. Shut pointless purposes or processes to scale back the load on the system.
   - **Replace Drivers and Software program**: Replace your working system, drivers (particularly graphics drivers), and every other vital software program to the most recent variations.
   - **Test for Malware or Viruses**: Run a full system scan utilizing a good antivirus program to detect and take away any malware.
   - **Regulate Energy Settings**: Regulate your energy settings to "Balanced" or "Energy Saver" mode (Home windows: Management Panel > Energy Choices, macOS: System Preferences > Vitality Saver).
   - **Examine the Laptop computer's {Hardware}**: If the laptop computer continues to be overheating, there could also be a difficulty with the {hardware}, corresponding to a failing fan or thermal paste that wants alternative. Seek the advice of knowledgeable technician to examine and exchange the thermal paste or defective {hardware} parts if vital.
   - **Environmental Components**: Function the laptop computer in a cool, well-ventilated atmosphere. Keep away from utilizing the laptop computer in direct daylight or close to warmth sources.
   - **Take into account Upgrading Elements**: If the laptop computer is older, contemplate upgrading parts corresponding to RAM or switching to an SSD to scale back the pressure on the system and assist with warmth administration.
   - **Monitor Temperature Ranges**: Set up a temperature monitoring instrument (e.g., HWMonitor, Core Temp, or Macs Fan Management) to maintain monitor of the CPU and GPU temperatures. This may help determine if the laptop computer is constantly operating at excessive temperatures and supply insights into when the overheating happens.

If the issue persists after attempting all of the above options, please contact the laptop computer producer’s help crew or an authorized restore technician for additional help.

**2. Duplicate Billing Problem:**
   - **Verification of Account Particulars**: I reviewed your account info to make sure accuracy and confirmed that the account is lively and matches the reported subject.  
   - **Billing Information Test**: Upon inspecting the billing data for final month, I recognized two equivalent fees processed on the identical day. This confirms the duplicate billing subject.  
   - **Cost Methodology Particulars**: I verified the fee methodology used for the transactions and confirmed that each fees have been processed utilizing the identical fee methodology.  
   - **Decision**:  
       - For the reason that duplicate cost was confirmed, I've initiated a refund for the second transaction. The refund will probably be processed again to the unique fee methodology inside 5-7 enterprise days.  
       - Moreover, I've utilized a credit score to your account as a gesture of goodwill for the inconvenience prompted.  
   - **Buyer Communication**: You'll obtain an e mail or telephone name from us to tell you of the decision, together with the refund timeline and the utilized credit score.  

We sincerely apologize for the inconvenience attributable to these points. Please tell us when you want additional help or have any further questions. We're right here to assist

And our ultimate, extra nuanced question is as follows:

Enter question: My bill quantity is inaccurate after system glitch?

The hint reveals that the question was appropriately categorised as ‘Each’, requiring responses from Technical and Billing brokers and a mixed response was supplied. Latency is 24 secs and tokens used are 12k.

Thanks for bringing this subject to our consideration. Primarily based on the technical investigation, listed here are the steps we're taking to resolve the system glitch that prompted the wrong bill quantity:

1. We have now verified the main points of the wrong bill quantity and recognized the difficulty within the system.
2. Our technical crew has analyzed the system logs and reviewed the bill calculation logic to pinpoint the foundation reason for the error.
3. The mandatory fixes have been applied to handle the glitch, and the system has been completely examined to make sure the difficulty is resolved.
4. A corrected bill has been generated, and we have now verified its accuracy.

Moreover, right here’s the entire decision for the billing subject associated to the wrong bill quantity attributable to the system glitch:

1. **Corrected Bill Verification**: The technical crew has resolved the system glitch and generated a corrected bill. The corrected bill has been personally verified to make sure its accuracy, together with cross-checking the quantities, dates, and any relevant fees or credit.

2. **Buyer Communication**: We are going to promptly notify you in regards to the decision of the difficulty. This communication contains:
   - An apology for the inconvenience attributable to the wrong bill.
   - Affirmation that the system glitch has been resolved.
   - Assurance that the corrected bill has been completely reviewed for accuracy.
   - A replica of the corrected bill in your data.

3. **Further Steps Taken**: To stop comparable points sooner or later, the technical crew has applied measures to make sure system stability and accuracy in bill era.

4. **Account Adjustment (if relevant)**: If the wrong bill resulted in any overpayment or underpayment, the mandatory changes will probably be made to your account. This contains issuing a refund for any overpayment or offering clear directions for settling any excellent steadiness.

5. **Comply with-Up**: We're right here to help you with any additional questions or issues concerning your account or billing. Please don't hesitate to succeed in out to us, and we will probably be completely happy to assist. On your comfort, we have now supplied direct contact info for additional communication.

We sincerely apologize for any inconvenience this will likely have prompted and guarantee you that we're taking steps to forestall comparable points sooner or later. Thanks in your understanding and persistence.

Analysis of this method

• High quality of orchestration: Good
• High quality of ultimate output: Good
• Explainability: Good (we perceive why it did what it did)
• Latency and Utilization: Honest (commensurate with the complexity of the output)

Takeaway

In abstract, the hierarchical Supervisor–Employee sample in CrewAI doesn’t perform as documented. The core orchestration logic is weak; as an alternative of permitting the supervisor to selectively delegate duties, CrewAI executes all duties sequentially, inflicting incorrect agent invocation, overwritten outputs, and inflated latency/token utilization. Why it failed comes all the way down to the framework’s inside routing—hierarchical mode doesn’t implement conditional branching or true delegation, so the ultimate response is successfully decided by whichever job occurs to run final. The repair is introducing a customized supervisor agent with specific, step-wise directions: it makes use of the triage end result, conditionally calls solely the required brokers, synthesizes their outputs, and terminates execution on the proper level—restoring right routing, enhancing output high quality, and considerably optimising token prices.

Conclusion

CrewAI, within the spirit of conserving the LLM on the middle of orchestration, relies upon upon it for many of the heavy-lifting of orchestration, utilising person prompts mixed with detailed scaffolding prompts embedded within the framework. In contrast to LangGraph and AutoGen, this method sacrifices determinism for developer-friendliness. And typically ends in surprising conduct for vital options such because the manager-worker sample, essential for a lot of real-life use circumstances. This text makes an attempt to exhibit a pathway for reaching the specified orchestration for this sample utilizing cautious prompting. In future articles, I intend to discover extra options for CrewAI, LangGraph and others for his or her applicability in sensible use circumstances.

You should utilize CrewAI to design an interactive conversational assistant on a doc retailer and additional make the responses really multimodal. Refer my articles on GraphRAG Design and Multimodal RAG.

Join with me and share your feedback at www.linkedin.com/in/partha-sarkar-lets-talk-AI

_{All photos on this article drawn by me or generated utilizing Copilot or Langfuse. Code shared is written by me.}