How do you pass context or outputs between tasks?

What you need to know

The default is "everything before me"

This surprises many people. In a sequential crew:

So task 5 with no context receives tasks 1 to 4 in full, on every step of its loop.

The four mechanisms

1research = Task(description="Research {company}.",2                expected_output="10 facts with URLs", agent=researcher)3profile = Task(description="Turn the research into a profile.",4               expected_output="A CompanyProfile",5               agent=analyst, context=[research],6               output_pydantic=CompanyProfile)7email = Task(description="Write a first email to {contact_name}.",8             expected_output="An email under 120 words",9             agent=writer, context=[profile])   # not the raw research1011crew.kickoff(inputs={"company": "Acme Logistics", "contact_name": "Priya"})

• context — email sees only the clean profile.
• Inputs — {company} and {contact_name} come from kickoff.
• Structured output — CompanyProfile is passed as JSON text, so the writer sees clear fields.
• Async join — if research had async_execution=True, context=[research] would make profile wait for it.

Passing data in a Flow

In a Flow, each step writes to self.state, and later steps read it. A common pattern: a crew step stores result.pydantic in state, and the next step passes selected fields as inputs to another crew. This gives you full control over what each crew sees.

A real-life example

A loan-document review crew has five tasks: extract salary slip, extract bank statement, extract ID, check policy, write summary. The summary task had no context, so it received all four earlier outputs — about 9,000 tokens, including raw OCR text — on each of its 5 steps. It sometimes quoted OCR errors ("Rs 8,50,00" instead of "Rs 85,000") that the policy task had already corrected.

The fix: the policy task outputs a PolicyResult model, and the summary task uses context=[policy_task]. The summary's input fell to about 1,200 tokens per step, cost per file dropped by about 40%, and OCR errors stopped leaking into credit notes.

Follow-up questions to expect

• "What is the difference between context and memory?" — Context is explicit and exact for this run; memory is retrieved by similarity, may include older runs, and is not guaranteed to surface a given fact.
• "Does context affect order?" — In a sequential crew the list order still decides; context makes a task wait for async tasks it depends on, but it cannot point to a task that comes later.
• "How do I pass a value only code knows, like a user ID?" — Use kickoff(inputs=...), or in a Flow keep it in state and pass it when calling the crew.