Build a parallel research agent in Python

This guide builds one deep-research action in ordinary asynchronous Python. The agent asks a model to plan a topic, investigates several briefs with bounded concurrency, and streams a synthesized report to its caller.

The implementation uses A11 for the model boundary, action nesting, output streams, logs, and remote serving. Python retains direct control of task creation, prompts, parsing, and application policy.

Define the agent's public boundary

The caller supplies one topic and can read the plan and report as each is produced. The model headers from interact_with_llm let a caller choose Claude, Gemini, or Ollama once for the whole run.

import asyncio

import a11
from a11.sdk.interact_with_llm import (
    INTERACT_WITH_LLM_SCHEMA,
    interact_with_llm,
)
from a11.sdk.llm import Interaction, Role


DEEP_RESEARCH_SCHEMA = a11.ActionSchema(
    name="deep-research",
    description=(
        "Plan a topic, investigate its parts concurrently, and write a report."
    ),
    inputs={
        "topic": a11.ActionPortSchema(
            "topic", "text/plain", typeinfo=str, unary=True, required=True
        )
    },
    outputs={
        "plan": a11.ActionPortSchema(
            "plan", "text/plain", typeinfo=str, required=True
        ),
        "report": a11.ActionPortSchema(
            "report", "text/plain", typeinfo=str, required=True
        ),
    },
    headers=INTERACT_WITH_LLM_SCHEMA.headers,
)

This schema is the product boundary. A browser or another service only needs the action name and ports; it does not need to know how the research is orchestrated.

Run model calls as nested actions

The helper below starts the included provider-neutral model action beneath the research action. propagate_io=False keeps its raw events, reasoning, and completed interactions in the process. Only text copied to a parent output is sent to the research caller.

The helper reads only the text it needs. When the parent action finishes in a session, A11 cleans up the unused nested outputs with the rest of the action.

async def _read_text(
    node: a11.AsyncNode,
    destination: a11.AsyncNode | None = None,
) -> str:
    parts: list[str] = []
    async for value in node:
        text = str(value)
        parts.append(text)
        if destination is not None:
            await destination.put(text)
    return "".join(parts)


async def ask_model(
    parent: a11.Action,
    prompt: str,
    *,
    stream_to: a11.AsyncNode | None = None,
) -> str:
    model = parent.make_nested(
        "interact_with_llm",
        propagate_io=False,
    ).run()

    text = asyncio.create_task(
        _read_text(model["text_output"], stream_to)
    )

    turn = Interaction(
        role=Role.USER,
        content=[a11.to_chunk({
            "role": "user",
            "content": [{"type": "text", "text": prompt}],
        })],
    )
    await model["interactions"].finalize(turn)
    await model["config"].finalize({})
    await model["tools"].finalize()

    answer = await text
    await model.wait()
    return answer

make_nested forwards the parent's A11 headers by default. The provider, model, credentials, deadline, and tracing context set on deep-research therefore reach every model call without becoming global state.

Orchestrate planning, investigation, and synthesis

The handler is a regular coroutine. A semaphore limits model concurrency, and asyncio.gather makes the barrier before synthesis explicit.

async def deep_research(action: a11.Action) -> None:
    topic = await action["topic"].consume(str)
    await action.logf("planning research on: %s", topic)

    plan_text = await ask_model(action, f"""
Plan research on this topic: {topic}

Write one independent investigation brief per line. Write at most three.
End with a line beginning `FINALLY:` that tells a writer how to synthesize
the findings. Do not number the lines.
""")
    lines = [line.strip() for line in plan_text.splitlines() if line.strip()]
    if len(lines) < 2 or not lines[-1].startswith("FINALLY:"):
        raise ValueError("the planner did not return briefs and a final task")

    briefs = lines[:-1]
    synthesis_brief = lines[-1].removeprefix("FINALLY:").strip()
    for brief in briefs:
        await action["plan"].put(brief)
    await action["plan"].finalize()

    limit = asyncio.Semaphore(3)

    async def investigate(brief: str) -> str:
        async with limit:
            await action.logf("investigating: %s", brief)
            return await ask_model(action, f"""
Research this brief in the context of `{topic}`:

{brief}

Return concise findings and name the sources you relied on.
""")

    findings = await asyncio.gather(
        *(investigate(brief) for brief in briefs)
    )

    await action.logf("synthesizing %d investigations", len(findings))
    joined_findings = "\n\n---\n\n".join(findings)
    await ask_model(
        action,
        f"""
Write the final report about `{topic}`.

Instruction: {synthesis_brief}

Investigation findings:
{joined_findings}
""",
        stream_to=action["report"],
    )
    await action["report"].finalize()

The planner completes before investigations start. Investigations overlap up to the explicit limit, and synthesis waits for their returned strings. The final model call writes tokens directly to the public report node, so the caller can render the report before the model finishes it.

If any child action or Python task fails, the parent action finishes with a non-OK status and its open outputs are aborted. Callers do not receive a normal end-of-stream for a partial report.

Register and serve the agent

Register the included model action and the Python research handler in the same registry:

registry = a11.ActionRegistry()
registry.register(
    "interact_with_llm",
    INTERACT_WITH_LLM_SCHEMA,
    interact_with_llm,
)
registry.register(
    "deep-research",
    DEEP_RESEARCH_SCHEMA,
    deep_research,
)

service = a11.Service(action_registry=registry)

The service can run over WebSocket, HTTP SSE, WebRTC, or another WireStream. The research handler stays unchanged when it moves behind a service or when a browser becomes its caller.

Try the deployed agent

Give the hosted demo a topic. The plan appears while the agent is preparing its investigations, activity follows the standard action log, and the report streams as it is synthesized. Several [investigate] entries overlap, while their full intermediate reports stay on the backend.

The hosted demo currently uses the optional Flow spelling described below. It exposes the same deep-research action boundary as the Python handler.

Running the demo backend

The default endpoint uses the hosted demo resources; bring your own key for a different provider. To run the backend locally:

python -m a11.demos.web_demos_server

This listens at ws://127.0.0.1:9010/a11-demos. A page loaded over HTTPS may block a plaintext WebSocket. Supply --certificate and --private-key with a trusted development certificate when needed.

A run makes one model call to plan, one per brief, and one to synthesize.

Provider Ollama Claude Gemini Model API key Base URL

Research

report port

Call it from a browser

The browser declares the same public ports and does not depend on the Python implementation:

const DEEP_RESEARCH_SCHEMA = new ActionSchema({
    name: 'deep-research',
    inputs: {
        topic: new ActionPortSchema({
            name: 'topic', type: 'text/plain', unary: true, required: true,
        }),
    },
    outputs: {
        report: new ActionPortSchema({name: 'report', type: 'text/plain'}),
        plan: new ActionPortSchema({name: 'plan', type: 'text/plain'}),
    },
});

Claim the log before dispatch, then read the plan and report concurrently:

const logs = await claimLog(call);
need(await call.call());

await Promise.all([
    readLogFrom(logs, (line) => addLine(log, line)),
    readPort(call, 'plan', (value) => addPlanItem(String(value))),
    readPort(call, 'report', (value) => appendReport(String(value))),
]);
need(await call.wait(600_000));

Reading all outputs concurrently prevents one bounded stream from stalling the others.

Optional: express the composition with Flow

Flow can describe the same orchestration when the composition should be loaded, checked, or changed at runtime. The corresponding section is concise because actions and streams provide the same execution model:

planned = run plan-research(topic: topic)
planned.briefs -> plan

nodes research {
  findings = node()

  for brief in planned.briefs parallel 3 {
    one = run investigate(topic: topic, brief: brief) timeout 2m
    one.report -> findings
  }

  written = run synthesise-findings(
    topic: topic, brief: planned.synthesis, findings: findings,
  )
  written.report -> report
}

Flow also makes the orchestration policies part of the loaded composition:

• The runtime drains every declared action output that the flow does not read. The Python illustration can leave unused nested outputs to session cleanup; Flow keeps them moving during the run. skip output records an intentional discard when that is useful to a reader.
• parallel 3 starts up to three investigations without task creation, semaphore, and gather code. The bound is part of the composition and applies when the source is loaded at runtime.
• timeout 2m limits each investigation step. Nested actions also inherit the parent's A11 headers, including its overall deadline, so one caller-supplied budget reaches planning, investigation, and synthesis.
• An unhandled action or stream failure ends the flow with the same structured status. The runtime stops outstanding steps and ends flow-owned streams, so the caller does not mistake partial output for success. cancel step can stop a named action when the composition implements a race or user abort.

Partial research needs an explicit recovery policy. try run investigate(...) allows the flow to inspect the step's status and continue; without try, the failure propagates. The nodes research block also keeps investigation ports and fragments on the service instead of sending them to the caller.

The Python handler is appropriate when orchestration is application code and normal Python control flow is the clearest expression. Flow is useful when the composition itself is runtime data, such as a checked plan supplied by a user or model, and it provides these lifecycle rules without more orchestration code. The deployed demo's complete Flow source is a11/demos/deep_research.flow.