Parallel Deep Research - Cited Reports With the Task API

The Task API runs deep, multi-step research and returns an answer with a "basis": the citations and confidence behind the findings. That is the difference between an answer and an answer you can verify.

deep_research.py

"""
Parallel Deep Research - Cited Reports With the Task API
========================================================

The Task API runs deep, multi-step research and returns an answer with a
"basis": the citations and confidence behind the findings. That is the
difference between an answer and an answer you can verify.

The agent calls create_task() to launch the research, then get_task_result()
to retrieve the report plus its sources.

Processors trade depth for time:
- "base"  - fast, good for most questions (seconds to a few minutes)
- "pro"   - deeper, and required for the "auto" output schema
- "ultra" - maximum depth (can run many minutes)

Prerequisites:
- pip install parallel-web
- export PARALLEL_API_KEY=<your-api-key>
"""

from agno.agent import Agent
from agno.models.openai import OpenAIResponses
from agno.tools.parallel import ParallelTools

# ---------------------------------------------------------------------------
# Tools - Task API (deep research)
# ---------------------------------------------------------------------------
# A "text" output schema returns a long-form markdown report with inline
# citations. Start with the base processor for a fast first pass.
research_tools = ParallelTools(
    enable_search=False,
    enable_extract=False,
    enable_task=True,
    default_processor="base",
    default_output_schema={"type": "text"},
)

# ---------------------------------------------------------------------------
# Create the Agent
# ---------------------------------------------------------------------------
research_agent = Agent(
    model=OpenAIResponses(id="gpt-5.4"),
    tools=[research_tools],
    markdown=True,
    instructions=[
        "Use create_task() to launch deep research, then get_task_result().",
        "Present the findings and list the sources behind each claim.",
    ],
)

# ---------------------------------------------------------------------------
# Run the Agent
# ---------------------------------------------------------------------------
if __name__ == "__main__":
    research_agent.print_response(
        "Research the current AI web-research API market: who the main "
        "providers are, how they price, and how they differ. Cite sources.",
        stream=True,
    )

Task completion

create_task returns a run ID; get_task_result waits for that run, up to default_timeout (1,800 seconds by default). Model streaming does not turn the research task into a streamed task result. Retain the run ID if you need to check a pending task later, and inspect an error result before treating it as completed research.

The processor list in the source is illustrative. Consult Parallel's current processor guide for available tiers and workload-dependent latency.

Run the Example

Set up your virtual environment

uv venv --python 3.12
source .venv/bin/activate

Install dependencies

uv pip install -U agno openai parallel-web

Export your API keys

export OPENAI_API_KEY="your_openai_api_key_here"
export PARALLEL_API_KEY="your_parallel_api_key_here"

Run the example

Save the code above as deep_research.py, then run:

python deep_research.py

Full source: cookbook/integrations/parallel/03_deep_research.py