Multi Agent Setup for Office Workers

In a previous post, we discussed how a developer sets up multi agents to increase productivity. In this post, we take a look at how office workers can take advantage of multi agent setups for non-coding tasks.

This New York Times article describes how white collar workers are benefiting from multi agent setups, but it doesn’t go into details on exactly how to do it. Here we give step by step instructions on how a non-coder can create multi agents for office tasks.

Anthropic

For Anthropic users, the options are limited. Anthropic doesn’t currently have a polished, no-code multi-agent product for non-technical users. One option is through Projects.

• Create separate Projects in Claude.ai for different functions (e.g. one for research, one for drafting, one for data analysis)
• Each project has its own persistent instructions and uploaded documents
• You manually carry outputs between them

Example for a business analyst:

Project: "Market Research"
  Instructions: "You are a research analyst. Always cite sources.
                 Summarize findings in bullet points."
  Uploaded docs: competitor reports, industry data

Project: "Executive Comms"
  Instructions: "You are a senior communications advisor.
                 Write in clear, concise business language."

Workflow: Research in Project 1 → copy findings → paste into Project 2 to draft the memo

The problem is that this doesn’t feel like a true multi agent set up because one has to manually communicate between the findings of each project.

OpenAI

OpenAI is slightly better for non-coders. ChatGPT Agent mode handles web tasks well, but it can’t access your local files, manage your calendar, send Slack messages, check your email, or take actions across the rest of your digital life. OpenAI has real multi-agent support, but it’s primarily a developer tool.

Google Gemini

Gemini, on the other hand, has the strongest support for web and Google Workspace support which is suitable for office tasks if your office uses Google Workspace.

The Three-Way Comparison for Non-Coders

	Claude	OpenAI	Google
Specialized assistants	Projects	Workspace Agents	Gemini Spark / Agent Studio
Scheduled/background tasks	Limited	Limited	Spark runs 24/7, even device off
Local file access	Cowork (desktop)	Weak	Google Drive native
Web task automation	Weaker	Agent Mode strong	Spark strong
Enterprise app integrations	MCP (setup required)	Workspace Agents	Native Google Workspace
Persistent memory	Strong (Projects)	Weak	Strong (Spark learns your style)
No-code agent builder	None native	Workspace Agents	Agent Studio visual
Availability	Now	Now	Spark still rolling out

A Realistic Day-in-the-Life Workflow

Say you’re a sales manager at a company on Google Workspace Enterprise:

Morning

• Ask Gemini in Chat gives you a daily briefing: unread threads, urgent action items, meetings today
• Meeting notes from yesterday’s call are already in Drive, summarized with action items

Mid-morning

• A new client inquiry arrives in Gmail
• Your Workspace Studio skill automatically extracts the lead details, logs them in your CRM Sheets tracker, and drafts a reply in your voice — you just review and send

Afternoon

• You need a proposal. Type: “Draft a proposal for [client] based on our last three similar proposals in Drive and their inquiry email”
• Gemini pulls the files, reads the email, and produces a formatted draft in Docs matching your company template

End of week

• A skill you built runs automatically every Friday: compiles the week’s deal updates from Sheets, meeting notes from Drive, and emails into a single status report, drops it in the team Chat channel

You can get sequential agent coordination through chaining other skills.

New email arrives
  → extract lead details        (step 1)
  → log to Sheets               (step 2)
  → draft reply                 (step 3)

Each step waits for the previous one. For true multi-agent parallel execution, you’ll need Gemini Enterprise which has an Agent Designer.

Gemini Enterprise Agent Designer is a purely visual and natural-language “No-Code” interface to create specialized subagents. You can create one Main Agent (The Orchestrator) that interacts with you. Inside that agent, you can build Subagents (The Workers). When you give the main agent a task, it will automatically break it down and assign different parts to the subagents.

OpenClaw

If you want to use an open source tool instead of relying on Gemini Enterprise, here’s how you can set up multi agents in OpenClaw.

The Concept: Multi-Agent Spaces

In OpenClaw, every agent is treated like an isolated “brain” with its own personality and folder. They run continuously on a background heartbeat loop. To make them work as a multi-agent team, you give them distinct roles using Markdown files, and then you link them to your messaging apps.

Step 1: Open Your OpenClaw Folder

When you install OpenClaw on your machine or server, it automatically creates a hidden folder in your user directory. You don’t need a code editor; just use your standard file explorer (Finder on Mac, File Explorer on Windows).

1. Go to your user folder and look for .openclaw (You may need to enable “Show Hidden Files” in your OS settings).
2. Inside, you will see an agents folder. This is where your team lives.
3. By default, there is a main folder. To create a new agent, simply copy the main folder and rename it (e.g., name one Researcher and another Writer).

Step 2: “Code” Your Agents Using Plain English (Markdown)

OpenClaw is designed to read text instructions. Inside each agent’s folder, you will find three vital Markdown files. You can open and edit these using any standard text editor (like Notepad or TextEdit).

You define your multi-agent roles by editing these files:

• SOUL.md: This is where you write the core identity and personality of the agent.
  • Example for your Researcher agent: "You are a meticulous, data-driven researcher. Your sole focus is analyzing web sources, verifying facts, and delivering concise summaries."
• AGENTS.md: This is where you tell the agent about the other agents in its ecosystem so it knows who to pass work to.
  • Example: "You are part of a team. Once you finish researching a topic, you must hand the summary over to the 'Writer' agent."
• USER.md: This tells the agent about you, your preferences, and your specific goals.

Step 3: Set Up Your Messaging Channels

Because OpenClaw doesn’t have a standard UI dashboard, your messaging apps act as the control center. You will want to create a separate bot account for each agent.

• For Telegram: Message the @BotFather on Telegram to create two separate bots (e.g., MyResearcherBot and MyWriterBot). Copy the HTTP API token for each.
• For Discord: Go to the Discord Developer Portal, create two separate applications/bots, turn on “Message Content Intent,” and copy their tokens.

Step 4: Route Your Agents (The openclaw.json file)

To tell the OpenClaw system which bot token goes to which Markdown folder, you will use the main configuration file located at ~/.openclaw/openclaw.json.

Even though it’s a JSON file, you can think of it like a simple form fill. Open it in a text editor and look for the agents and bindings sections. You will paste your bot tokens and map them to the folders you created:

{
  "agents": {
    "list": [
      { "agentId": "researcher", "agentDir": "~/.openclaw/agents/researcher" },
      { "agentId": "writer", "agentDir": "~/.openclaw/agents/writer" }
    ]
  },
  "channels": {
    "telegram": {
      "accounts": {
        "researcher_bot": { "token": "PASTE_YOUR_RESEARCHER_TOKEN_HERE" },
        "writer_bot": { "token": "PASTE_YOUR_WRITER_TOKEN_HERE" }
      },
      "bindings": [
        { "accountId": "researcher_bot", "agentId": "researcher" },
        { "accountId": "writer_bot", "agentId": "writer" }
      ]
    }
  }
}

Step 5: Put Them to Work

Once you launch OpenClaw, both agents are officially alive and running in the background.

1. Direct Message them: You can chat with your Researcher bot on Telegram to give it a task. It will use its local tools to browse the web or read your files.
2. Group Chats (The Collaboration): Create a Telegram group or a Discord server and invite both bots to it.
3. Because you defined their relationship in AGENTS.md, you can tag the researcher in a group chat and say: “@Researcher, find the top 3 trends in AI for 2026.”
4. Once the researcher outputs the data, the writer agent (noticing its queue) will automatically trigger its heartbeat cycle, read the researcher’s output, and say: “Taking it from here. Designing the newsletter layout based on those 3 trends…”