CrewAI Task Never Completes: Diagnosing the Silent Hang

Nothing is hung. One agent is stuck, and everything downstream is politely waiting for it. Find the one whose tool results stopped changing and you have found your problem.

Find the stuck agent

CrewAI's default output tells you which task is running but not what the agent inside it is doing. Add a step callback that records the tool name, a timestamp and a hash of the result. The stuck agent is the one producing the same hash repeatedly, or the one whose last call never returned.

import hashlib, time

def step(output):
    digest = hashlib.sha256(str(output).encode()).hexdigest()[:10]
    print(f"{time.strftime('%H:%M:%S')} {output.tool} -> {digest}")

crew = Crew(agents=[...], tasks=[...], step_callback=step)

The usual culprits

• A browser tool on a login page, retrying into it and getting the same page back.
• A rate-limited API returning the same error the agent keeps re-reading as content.
• A page that never finished loading, so every scrape returns the loading skeleton.

Give the tool a way out

The fix is in the tool, not the crew. Any tool that touches the outside world should be able to say I am blocked and hand the situation to a person, rather than returning the same result forever. Two tools do it: one that asks whether a human has already solved this wall, and one that asks a human now. Give the browsing agent both, with descriptions that tell the model to check first and ask second.

With those in place a stuck crew stops being a twenty-minute mystery. The agent asks, a person clears the wall in the live browser, the tool returns real output, and the rest of the crew proceeds. The next crew that hits the same wall gets the answer from the first call.

Questions people ask