> For the complete documentation index, see [llms.txt](https://docs.midio.com/midio-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.midio.com/midio-docs/tutorials/concurrent-research-agent.md). # Concurrent Research Agent In this guide we will create a **Research Agent** that takes a given user question and automatically does several things: * Breaks the question into discrete research tasks * Launches multiple sub‑agents that search the web in parallel * Collects and deduplicates their findings * Produces a concise, well‑structured final answer We leverage the built‑in concurrency primitives such as **`Std.Spawn Processes from List`** and **`Std.MergeAll`**. {% embed url="" %} *** ## High‑Level Architecture Our solution is split into three main phases: | Phase | Responsibility | | ------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **1) Task Planning** | Turn the user prompt into a list of concrete research tasks | | **2) Concurrent Research** |

Spawn one child process per task, each of which:
• Transforms its task into a search engine query
• Queries the Tavily Search API
• Summarizes the results

| | **3) Aggregation & Synthesis** | Merge all child outputs, filter failures, and craft the master answer |

## Phase 1: Task Planning The master agent needs a **machine‑readable** list of subtasks. Midio’s `Chat Complete` (avilable in the open-ai [package](/midio-docs/package-manager.md)) node supports [**structured output**](/midio-docs/guides/building-agents.md#structured-output) which lets an LLM return valid JSON that is guaranteed to match a user‑defined schema. 1. We can force the LLM to output a custom schema by providing an object like the following to the `response format` input. We also need to mention in the system prompt that we expect JSON output. ```json { name: "AgentDeployments", description: "A list of agents to deploy, which agents and what their task is.", schema: { agents: [ { name: "WebResearchAgent", task: "A description of what the agent should research." } ] } } ``` 2. To the `system message` input, we pass the following ``` You are a senior researcher. Break the following question into distinct, non‑overlapping web research tasks. Spawn at least 3 tasks, at most 6. Return an array of objects that conforms to the provided JSON Schema. ``` 3. And the users query is passed directly to the `user message` input.

## Phase 2: Concurrent Research ### 1. Spawning Child Processes We can spawn a process for each generated task by using the [**`Std.Spawn Processes from List`**](/midio-docs/built-in-nodes/core-std.md#spawn-processes-from-list) node, and connect the agents field from the response of the `Generate Research Tasks` function. We can then connect the `child spawned` trigger (which is triggered once per item) to the `Search Web and Summarize` function. ### Inside the Research Sub‑Agent The sub‑agent receives a single task object. It first searches the web, and then summarizes its findings. Here is an example of a task object. Generated by the previous step. ```json { "task": "Compare the energy density of li‑ion vs. sodium‑ion batteries" } ``` ### 2. Craft a Search Query A `Chat Complete` call turns the (sometimes verbose) task into a concise search‑engine query using the following prompt: {% code overflow="wrap" %} ``` You are given a task, and you should convert it into a search query that you might put into a google search, and expect resources that provide information to answer the task. ``` {% endcode %} The response is then piped directly into the query input of a [`Tavily.Search`](http://tavily.search) node. You can add the Tavily package using the package manager in the editor.

### 3. Convert Pages to Markdown For each search result we get back from Tavily, we convert it to Markdown using the `Html to Markdown` function in the **web-tools** [package](/midio-docs/package-manager.md). We use `Spawn Processes from List` here as well to do this task concurrently. {% hint style="info" %} Tavily supports returning a summary of the page results automatically, but we do it manually here for demo purposes. {% endhint %}

### 4. Summarize Findings Another `Chat Complete` node receives both the **task** and **raw search results**, and is prompted to combine them into a task summary. {% code overflow="wrap" %} ``` You are a web research summarizer agent. The user will provide you the result of a web search, as well a description of a research task that you are tasked with addressing. Use the web search results and task description to create a summary. Expect the output of your summary to be used by another agent which will combine it with other search results performed by other agents similar to you. ``` {% endcode %} ## Phase 3: Aggregation & Synthesis The final step is to combine all the research data and produce a single response. We do this with another LLM call, with the following prompt: {% code overflow="wrap" %} ``` You are a web research summarizer agent. The user will provide you the result of a web search, as well a description of a research task that you are tasked with addressing. Use the web search results and task description to create a summary. Expect the output of your summary to be used by another agent which will combine it with other search results performed by other agents similar to you. ``` {% endcode %} For its user message we pass the search results and the original user query using the following expression string template: ``` " ${data} ${userQuery} " ```

{% hint style="info" %} We also make sure to remove any null values from the inputs list, which might appear if a research agent fails for some reason. {% endhint %} ### Exposing the Agent as a REST API To make the agent callable from any frontend (e.g. Lovable, curl, or your own app), we can hook it up to an `Endpoint` event node like this:

And return a response like this:

Your research micro service is now live at: [`https://.midio.dev:3001/research`](https://github.com/midio-code/midio-docs/blob/master/tutorials/https:/%3Cyour-project-name%3E.midio.dev:3001/research?query/README.md) and expecting a single query parameter named `query`. You can call it like this using curl, for example: {% code overflow="wrap" %} ``` curl https://.midio.dev:3001/research?query=what+is+going+on+in+tech ``` {% endcode %}