Parallel workflow

Parallelization takes advantage of tasks that can be broken up into discrete independent parts. The user’s prompt is passed to multiple LLMs simultaneously. Once all the LLMs respond, their answers are all sent to a final LLM call to be aggregated for the final answer.

Parallel Architecture

Run multiple LLMs in parallel and aggregate their solutions.

Notice that the same user prompt goes to each parallel LLM for execution. An alternate parallel workflow where this main prompt task is broken into sub-tasks is presented later.

Parallel Workflow Cookbook

For a more detailed walk-through refer to the notebook here.

Setup Client & Helper Functions

import asyncio
import together
from together import AsyncTogether, Together

client = Together()
async_client = AsyncTogether()


def run_llm(user_prompt: str, model: str, system_prompt: str = None):
    messages = []
    if system_prompt:
        messages.append({"role": "system", "content": system_prompt})

    messages.append({"role": "user", "content": user_prompt})

    response = client.chat.completions.create(
        model=model,
        messages=messages,
        temperature=0.7,
        max_tokens=4000,
    )

    return response.choices[0].message.content


# The function below will call the reference LLMs in parallel
async def run_llm_parallel(
    user_prompt: str,
    model: str,
    system_prompt: str = None,
):
    """Run a single LLM call with a reference model."""
    for sleep_time in [1, 2, 4]:
        try:
            messages = []
            if system_prompt:
                messages.append({"role": "system", "content": system_prompt})

            messages.append({"role": "user", "content": user_prompt})

            response = await async_client.chat.completions.create(
                model=model,
                messages=messages,
                temperature=0.7,
                max_tokens=2000,
            )
            break
        except together.error.RateLimitError as e:
            print(e)
            await asyncio.sleep(sleep_time)
    return response.choices[0].message.content

import assert from "node:assert";
import Together from "together-ai";

const client = new Together();

export async function runLLM(
  userPrompt: string,
  model: string,
  systemPrompt?: string,
) {
  const messages: { role: "system" | "user"; content: string }[] = [];
  if (systemPrompt) {
    messages.push({ role: "system", content: systemPrompt });
  }

  messages.push({ role: "user", content: userPrompt });

  const response = await client.chat.completions.create({
    model,
    messages,
    temperature: 0.7,
    max_tokens: 4000,
  });

  const content = response.choices[0].message?.content;
  assert(typeof content === "string");
  return content;
}

Implement Workflow

import asyncio
from typing import List


async def parallel_workflow(
    prompt: str,
    proposer_models: List[str],
    aggregator_model: str,
    aggregator_prompt: str,
):
    """Run a parallel chain of LLM calls to address the `input_query`
    using a list of models specified in `models`.

    Returns output from final aggregator model.
    """

    # Gather intermediate responses from proposer models
    proposed_responses = await asyncio.gather(
        *[run_llm_parallel(prompt, model) for model in proposer_models]
    )

    # Aggregate responses using an aggregator model
    final_output = run_llm(
        user_prompt=prompt,
        model=aggregator_model,
        system_prompt=aggregator_prompt
        + "\n"
        + "\n".join(
            f"{i+1}. {str(element)}"
            for i, element in enumerate(proposed_responses)
        ),
    )

    return final_output, proposed_responses

import dedent from "dedent";

/*
  Run a parallel chain of LLM calls to address the `inputQuery`
  using a list of models specified in `proposerModels`.

  Returns output from final aggregator model.
*/
async function parallelWorkflow(
  inputQuery: string,
  proposerModels: string[],
  aggregatorModel: string,
  aggregatorSystemPrompt: string,
) {
  // Gather intermediate responses from proposer models
  const proposedResponses = await Promise.all(
    proposerModels.map((model) => runLLM(inputQuery, model)),
  );

  // Aggregate responses using an aggregator model
  const aggregatorSystemPromptWithResponses = dedent`
    ${aggregatorSystemPrompt}

    ${proposedResponses.map((response, i) => `${i + 1}. response`)}
  `;

  const finalOutput = await runLLM(
    inputQuery,
    aggregatorModel,
    aggregatorSystemPromptWithResponses,
  );

  return [finalOutput, proposedResponses];
}

Example Usage

reference_models = [
    "zai-org/GLM-5.2",
    "openai/gpt-oss-120b",
    "meta-llama/Llama-3.3-70B-Instruct-Turbo",
    "MiniMaxAI/MiniMax-M2.7",
]

user_prompt = """Jenna and her mother picked some apples from their apple farm.
Jenna picked half as many apples as her mom. If her mom got 20 apples, how many apples did they both pick?"""

aggregator_model = "deepseek-ai/DeepSeek-V3"

aggregator_system_prompt = """You have been provided with a set of responses from various open-source models to the latest user query.
Your task is to synthesize these responses into a single, high-quality response. It is crucial to critically evaluate the information
provided in these responses, recognizing that some of it may be biased or incorrect. Your response should not simply replicate the
given answers but should offer a refined, accurate, and comprehensive reply to the instruction. Ensure your response is well-structured,
coherent, and adheres to the highest standards of accuracy and reliability.

Responses from models:"""


async def main():
    answer, intermediate_responses = await parallel_workflow(
        prompt=user_prompt,
        proposer_models=reference_models,
        aggregator_model=aggregator_model,
        aggregator_prompt=aggregator_system_prompt,
    )

    for i, response in enumerate(intermediate_responses):
        print(f"Intermediate Response {i+1}:\n\n{response}\n")

    print(f"Final Answer: {answer}\n")


asyncio.run(main())

const referenceModels = [
  "zai-org/GLM-5.2",
  "openai/gpt-oss-120b",
  "meta-llama/Llama-3.3-70B-Instruct-Turbo",
  "MiniMaxAI/MiniMax-M2.7",
];

const userPrompt = dedent`
  Jenna and her mother picked some apples from their apple farm.
  Jenna picked half as many apples as her mom.

  If her mom got 20 apples, how many apples did they both pick?
`;

const aggregatorModel = "deepseek-ai/DeepSeek-V3";

const aggregatorSystemPrompt = dedent`
  You have been provided with a set of responses from various
  open-source models to the latest user query. Your task is to
  synthesize these responses into a single, high-quality response.
  It is crucial to critically evaluate the information provided in
  these responses, recognizing that some of it may be biased or incorrect.
  Your response should not simply replicate the given answers but
  should offer a refined, accurate, and comprehensive reply to the
  instruction. Ensure your response is well-structured, coherent, and
  adheres to the highest standards of accuracy and reliability.

  Responses from models:
`;

async function main() {
  const [answer, intermediateResponses] = await parallelWorkflow(
    userPrompt,
    referenceModels,
    aggregatorModel,
    aggregatorSystemPrompt,
  );
  for (const response of intermediateResponses) {
    console.log(
      `## Intermediate Response: ${intermediateResponses.indexOf(response) + 1}:\n`,
    );
    console.log(`${response}\n`);
  }
  console.log(`## Final Answer:`);
  console.log(`${answer}\n`);
}

main();

Use cases

Using one LLM to answer a user’s question, while at the same time using another to screen the question for inappropriate content or requests.
Reviewing a piece of code for both security vulnerabilities and stylistic improvements at the same time.
Analyzing a lengthy document by dividing it into sections and assigning each section to a separate LLM for summarization, then combining the summaries into a comprehensive overview.
Simultaneously analyzing a text for emotional tone, intent, and potential biases, with each aspect handled by a dedicated LLM.
Translating a document into multiple languages at the same time by assigning each language to a separate LLM, then aggregating the results for multilingual output.

Subtask Agent Workflow

An alternate and useful parallel workflow. This workflow begins with an LLM breaking down the task into subtasks that are dynamically determined based on the input. These subtasks are then processed in parallel by multiple worker LLMs. Finally, the orchestrator LLM synthesizes the workers’ outputs into the final result.

Subtask Workflow Cookbook

For a more detailed walk-through refer to the notebook here.

Setup Client & Helper Functions

import asyncio
import json
import together
from pydantic import ValidationError
from together import AsyncTogether, Together

client = Together()
async_client = AsyncTogether()


# The function below will call the reference LLMs in parallel
async def run_llm_parallel(
    user_prompt: str,
    model: str,
    system_prompt: str = None,
):
    """Run a single LLM call with a reference model."""
    for sleep_time in [1, 2, 4]:
        try:
            messages = []
            if system_prompt:
                messages.append({"role": "system", "content": system_prompt})

            messages.append({"role": "user", "content": user_prompt})

            response = await async_client.chat.completions.create(
                model=model,
                messages=messages,
                temperature=0.7,
                max_tokens=2000,
            )
            break
        except together.error.RateLimitError as e:
            print(e)
            await asyncio.sleep(sleep_time)
    return response.choices[0].message.content


def JSON_llm(user_prompt: str, schema, system_prompt: str = None):
    try:
        messages = []
        if system_prompt:
            messages.append({"role": "system", "content": system_prompt})

        messages.append({"role": "user", "content": user_prompt})

        extract = client.chat.completions.create(
            messages=messages,
            model="meta-llama/Llama-3.3-70B-Instruct-Turbo",
            response_format={
                "type": "json_schema",
                "json_schema": {
                    "name": "response",
                    "schema": schema.model_json_schema(),
                },
            },
        )
        return json.loads(extract.choices[0].message.content)

    except ValidationError as e:
        error_message = f"Failed to parse JSON: {e}"
        print(error_message)

import assert from "node:assert";
import Together from "together-ai";
import { z, type ZodType } from "zod";

const client = new Together();

export async function runLLM(userPrompt: string, model: string) {
  const response = await client.chat.completions.create({
    model,
    messages: [{ role: "user", content: userPrompt }],
    temperature: 0.7,
    max_tokens: 4000,
  });

  const content = response.choices[0].message?.content;
  assert(typeof content === "string");
  return content;
}

export async function jsonLLM<T>(
  userPrompt: string,
  schema: ZodType<T>,
  systemPrompt?: string,
) {
  const messages: { role: "system" | "user"; content: string }[] = [];
  if (systemPrompt) {
    messages.push({ role: "system", content: systemPrompt });
  }

  messages.push({ role: "user", content: userPrompt });

  const response = await client.chat.completions.create({
    model: "meta-llama/Llama-3.3-70B-Instruct-Turbo",
    messages,
    response_format: {
      type: "json_schema",
      json_schema: {
        name: "response",
        schema: z.toJSONSchema(schema),
      },
    },
  });

  const content = response.choices[0].message?.content;
  assert(typeof content === "string");

  return schema.parse(JSON.parse(content));
}

Implement Workflow

import asyncio
import json
from pydantic import BaseModel, Field
from typing import Literal, List

ORCHESTRATOR_PROMPT = """
Analyze this task and break it down into 2-3 distinct approaches:

Task: {task}

Provide an Analysis:

Explain your understanding of the task and which variations would be valuable.
Focus on how each approach serves different aspects of the task.

Along with the analysis, provide 2-3 approaches to tackle the task, each with a brief description:

Formal style: Write technically and precisely, focusing on detailed specifications
Conversational style: Write in a friendly and engaging way that connects with the reader
Hybrid style: Tell a story that includes technical details, combining emotional elements with specifications

Return only JSON output.
"""

WORKER_PROMPT = """
Generate content based on:
Task: {original_task}
Style: {task_type}
Guidelines: {task_description}

Return only your response:
[Your content here, maintaining the specified style and fully addressing requirements.]
"""

task = """Write a product description for a new eco-friendly water bottle.
The target_audience is environmentally conscious millennials and key product features are: plastic-free, insulated, lifetime warranty
"""


class Task(BaseModel):
    type: Literal["formal", "conversational", "hybrid"]
    description: str


class TaskList(BaseModel):
    analysis: str
    tasks: List[Task] = Field(..., default_factory=list)


async def orchestrator_workflow(
    task: str,
    orchestrator_prompt: str,
    worker_prompt: str,
):
    """Use an orchestrator model to break down a task into sub-tasks and then use worker models to generate and return responses."""

    # Use orchestrator model to break the task up into sub-tasks
    orchestrator_response = JSON_llm(
        orchestrator_prompt.format(task=task),
        schema=TaskList,
    )

    # Parse orchestrator response
    analysis = orchestrator_response["analysis"]
    tasks = orchestrator_response["tasks"]

    print("\n=== ORCHESTRATOR OUTPUT ===")
    print(f"\nANALYSIS:\n{analysis}")
    print(f"\nTASKS:\n{json.dumps(tasks, indent=2)}")

    worker_model = ["meta-llama/Llama-3.3-70B-Instruct-Turbo"] * len(tasks)

    # Gather intermediate responses from worker models
    return tasks, await asyncio.gather(
        *[
            run_llm_parallel(
                user_prompt=worker_prompt.format(
                    original_task=task,
                    task_type=task_info["type"],
                    task_description=task_info["description"],
                ),
                model=model,
            )
            for task_info, model in zip(tasks, worker_model)
        ]
    )

import dedent from "dedent";
import { z } from "zod";

function ORCHESTRATOR_PROMPT(task: string) {
  return dedent`
    Analyze this task and break it down into 2-3 distinct approaches:

    Task: ${task}

    Provide an Analysis:

    Explain your understanding of the task and which variations would be valuable.
    Focus on how each approach serves different aspects of the task.

    Along with the analysis, provide 2-3 approaches to tackle the task, each with a brief description:

    Formal style: Write technically and precisely, focusing on detailed specifications
    Conversational style: Write in a friendly and engaging way that connects with the reader
    Hybrid style: Tell a story that includes technical details, combining emotional elements with specifications

    Return only JSON output.
  `;
}

function WORKER_PROMPT(
  originalTask: string,
  taskType: string,
  taskDescription: string,
) {
  return dedent`
    Generate content based on:
    Task: ${originalTask}
    Style: ${taskType}
    Guidelines: ${taskDescription}

    Return only your response:
    [Your content here, maintaining the specified style and fully addressing requirements.]
  `;
}

const taskListSchema = z.object({
  analysis: z.string(),
  tasks: z.array(
    z.object({
      type: z.enum(["formal", "conversational", "hybrid"]),
      description: z.string(),
    }),
  ),
});

/*
  Use an orchestrator model to break down a task into sub-tasks,
  then use worker models to generate and return responses.
*/
async function orchestratorWorkflow(
  originalTask: string,
  orchestratorPrompt: (task: string) => string,
  workerPrompt: (
    originalTask: string,
    taskType: string,
    taskDescription: string,
  ) => string,
) {
  // Use orchestrator model to break the task up into sub-tasks
  const { analysis, tasks } = await jsonLLM(
    orchestratorPrompt(originalTask),
    taskListSchema,
  );

  console.log(dedent`
    ## Analysis:
    ${analysis}

    ## Tasks:
  `);
  console.log("```json", JSON.stringify(tasks, null, 2), "\n```\n");

  const workerResponses = await Promise.all(
    tasks.map(async (task) => {
      const response = await runLLM(
        workerPrompt(originalTask, task.type, task.description),
        "meta-llama/Llama-3.3-70B-Instruct-Turbo",
      );

      return { task, response };
    }),
  );

  return workerResponses;
}

Example Usage

async function main() {
  const task = `Write a product description for a new eco-friendly water bottle.
    The target_audience is environmentally conscious millennials and key product
    features are: plastic-free, insulated, lifetime warranty
  `;

  const workerResponses = await orchestratorWorkflow(
    task,
    ORCHESTRATOR_PROMPT,
    WORKER_PROMPT,
  );

  console.log(
    workerResponses
      .map((w) => `## WORKER RESULT (${w.task.type})\n${w.response}`)
      .join("\n\n"),
  );
}

main();

async function main() {
  const task = `Write a product description for a new eco-friendly water bottle.
    The target_audience is environmentally conscious millennials and key product
    features are: plastic-free, insulated, lifetime warranty
  `;

  const workerResponses = await orchestratorWorkflow(
    task,
    ORCHESTRATOR_PROMPT,
    WORKER_PROMPT,
  );

  console.log(
    workerResponses
      .map((w) => `## WORKER RESULT (${w.task.type})\n${w.response}`)
      .join("\n\n"),
  );
}

main();

Use cases

Breaking down a coding problem into subtasks, using an LLM to generate code for each subtask, and making a final LLM call to combine the results into a complete solution.
Searching for data across multiple sources, using an LLM to identify relevant sources, and synthesizing the findings into a cohesive answer.
Creating a tutorial by splitting each section into subtasks like writing an introduction, outlining steps, and generating examples. Worker LLMs handle each part, and the orchestrator combines them into a polished final document.
Dividing a data analysis task into subtasks like cleaning the data, identifying trends, and generating visualizations. Each step is handled by separate worker LLMs, and the orchestrator integrates their findings into a complete analytical report.

GUIDES

MODEL QUICKSTARTS

BUILD APPS

BUILD AGENTS

WORK WITH CODING AGENTS

RAG & SEARCH

SDK INTEGRATIONS

DEDICATED CONTAINERS

Parallel Architecture

Parallel Workflow Cookbook

Setup Client & Helper Functions

Implement Workflow

Example Usage

Use cases

Subtask Agent Workflow

Subtask Workflow Cookbook

Setup Client & Helper Functions

Implement Workflow

Example Usage

Use cases

​Parallel Architecture

​Parallel Workflow Cookbook

​Setup Client & Helper Functions

​Implement Workflow

​Example Usage

​Use cases

​Subtask Agent Workflow

​Subtask Workflow Cookbook

​Setup Client & Helper Functions

​Implement Workflow

​Example Usage

​Use cases

Parallel Architecture

Parallel Workflow Cookbook

Setup Client & Helper Functions

Implement Workflow

Example Usage

Use cases

Subtask Agent Workflow

Subtask Workflow Cookbook

Setup Client & Helper Functions

Implement Workflow

Example Usage

Use cases