How to Fix 'JSON parsing error when scaling' in CrewAI (Python)

What this error means

JSON parsing error when scaling in CrewAI usually means one of your agents or tools returned text that CrewAI expected to be valid JSON, but it got extra prose, malformed structure, or a type it couldn’t serialize. You typically see this when using structured outputs, tool calls, or when an agent is asked to return data that later gets parsed during task execution.

In practice, this shows up during crew.kickoff(), especially when a Task expects structured output or a tool returns a Python object instead of a JSON-safe string.

The Most Common Cause

The #1 cause is an agent returning non-JSON text when the task expects JSON, or a tool returning Python objects that are not serializable.

A common mistake is telling the model to “return JSON” without enforcing a schema, then parsing the result as if it were guaranteed valid.

# BROKEN
from crewai import Agent, Task, Crew
import json

researcher = Agent(
    role="Researcher",
    goal="Return customer risk summary as JSON",
    backstory="You analyze insurance claims."
)

task = Task(
    description="Summarize the claim and return JSON only.",
    agent=researcher,
)

crew = Crew(agents=[researcher], tasks=[task])
result = crew.kickoff()

# This often fails because result is not strict JSON
data = json.loads(result.raw)

# FIXED
from pydantic import BaseModel
from crewai import Agent, Task, Crew

class ClaimSummary(BaseModel):
    claim_id: str
    risk_level: str
    reasons: list[str]

researcher = Agent(
    role="Researcher",
    goal="Return customer risk summary in the exact schema.",
    backstory="You analyze insurance claims.",
)

task = Task(
    description=(
        "Summarize the claim using only the required fields. "
        "Do not add markdown or explanation."
    ),
    agent=researcher,
    expected_output=ClaimSummary.model_json_schema(),
)

crew = Crew(agents=[researcher], tasks=[task])
result = crew.kickoff()

If you’re using tools, make sure they return strings that can be parsed cleanly:

# BROKEN TOOL OUTPUT
def get_claim_data(claim_id: str):
    return {"claim_id": claim_id, "amount": 12000}  # may break downstream parsing

# FIXED TOOL OUTPUT
import json

def get_claim_data(claim_id: str):
    return json.dumps({"claim_id": claim_id, "amount": 12000})

Other Possible Causes

1) The model wrapped JSON in markdown fences

CrewAI may receive something like:

{
  "claim_id": "C-1001"
}

That looks fine to humans, but json.loads() will fail if the code fences are included in the raw output.

# BAD PROMPT
"Return JSON only in a markdown block."

# BETTER PROMPT
"Return raw JSON only. No markdown fences. No explanation."

2) A tool returned None or an unsupported type

If a tool returns None, bytes, datetime objects, or custom classes, serialization can fail during task scaling or handoff.

# BROKEN
from datetime import datetime

def audit_tool():
    return {"checked_at": datetime.utcnow()}  # not JSON serializable

# FIXED
from datetime import datetime

def audit_tool():
    return {"checked_at": datetime.utcnow().isoformat()}

3) Nested quotes or trailing commas in generated JSON

LLMs often produce invalid JSON like trailing commas or unescaped quotes.

# BROKEN OUTPUT
{
  "name": "ACME "Gold" Plan",
  "status": "active",
}

# FIXED OUTPUT
{
  "name": "ACME Gold Plan",
  "status": "active"
}

If you need this reliably, don’t parse free-form text. Use schema validation after generation.

4) Version mismatch between CrewAI and dependencies

I’ve seen parsing issues caused by mismatched versions of crewai, pydantic, and provider SDKs. This is common after upgrading one package without pinning the rest.

crewai==0.x.x
pydantic==2.x.x
openai==1.x.x

Check for pinned versions and upgrade them together. If the issue started after an upgrade, roll back and test.

How to Debug It

1. •

   Inspect the raw output before parsing

   • •Log result.raw or the tool response exactly as returned.
   • •Look for markdown fences, extra commentary, trailing commas, or empty values.
2. •

   Reduce the task to one agent and one tool

   • •Remove delegation, memory, and multiple tools.
   • •If the error disappears, one of those layers is producing invalid data during scaling.
3. •

   Validate against a schema immediately

   • •Use Pydantic to catch bad output at the boundary.
   • •Example:

from pydantic import BaseModel, ValidationError

class OutputSchema(BaseModel):
    claim_id: str
    status: str

try:
    parsed = OutputSchema.model_validate_json(result.raw)
except ValidationError as e:
    print(e)

4. •Check whether the failure happens at kickoff or handoff
   • •If it fails during crew.kickoff(), inspect task outputs.
   • •If it fails during delegation/scaling across agents, inspect intermediate outputs from each agent and each tool call.

Prevention

• •

  Use typed outputs everywhere

  • •Prefer Pydantic models over “JSON-only” prompts.
  • •Validate at every boundary between agent output and Python code.
• •

  Keep tool outputs JSON-safe

  • •Return strings, numbers, booleans, lists, and dicts with primitive values only.
  • •Convert datetimes with .isoformat() and avoid custom objects.
• •

  Pin your dependency versions

  • •Lock crewai, pydantic, and provider SDKs together.
  • •Re-test after upgrades before deploying to production workflows.

If you’re seeing JSON parsing error when scaling in CrewAI Python code, start with the raw output first. In most cases, the bug is not “CrewAI scaling” itself — it’s invalid structured data slipping through an agent or tool boundary.

Keep learning

• •The complete AI Agents Roadmap — my full 8-step breakdown
• •Free: The AI Agent Starter Kit — PDF checklist + starter code
• •Work with me — I build AI for banks and insurance companies

By Cyprian Aarons, AI Consultant at Topiax.