My Multi-Agent Team with OpenClaw

Setting Up AI Agents with OpenClaw

Introduction to the Setup

• The speaker introduces a new Mac Mini dedicated to running a team of AI agents using OpenClaw, which includes roles like developer, marketer, project manager, and system admin.

• Initial setup was complex and required addressing various technical and strategic questions regarding hardware choices and software capabilities.

Key Considerations in Setup

• Important considerations included whether to use a VPS or a physical machine, costs associated with API tokens, chat tool preferences (Telegram vs. WhatsApp vs. Slack), and security measures for agent access.

• The speaker reflects on their initial skepticism about OpenClaw's utility but later recognizes its potential in solving business challenges.

Business Context for Using OpenClaw

• The speaker discusses their YouTube channel and course offerings while highlighting the bottleneck in content creation due to limited bandwidth.

• Previous solutions involved hiring real teams; however, the overhead costs prompted exploration of autonomous agents as an alternative.

Paradigm Shift with Autonomous Agents

• The realization that autonomous agents could fill specific roles within the team led to a shift in perspective about AI's application in business operations.

• Emphasizes that OpenClaw represents just the beginning of what could be a larger trend towards utilizing AI agents effectively.

Understanding OpenClaw's Functionality

Core Features of OpenClaw

• OpenClaw is described as having evolved from previous iterations (Cloudbot and Moltbot), focusing on its core component known as the gateway.

• Unlike traditional cloud code usage, OpenClaw maintains persistent workspaces allowing agents to operate independently through messaging platforms like Telegram or Slack.

Deployment Recommendations

• It is advised not to run OpenClaw on personal machines due to security risks; instead, it should have its own dedicated environment.

• Options include using a cloud VPS or any computer on the network; the speaker chose a Mac Mini M4 for better management visibility.

Security Measures for Agent Access

• The importance of restricting agent access is highlighted; similar to hiring practices where employees are given limited permissions based on their roles.

• A dedicated email address and GitHub username were created specifically for agent interactions while ensuring proper access controls are maintained.

Accessing Services and File Management

Setting Up OpenClaw Workspace

• The speaker discusses the need for easy two-way syncing between their computer and the OpenClaw workspace on a Mac Mini, emphasizing the development of a brain system that logs business activities into Markdown files accessible by agents.

• Files are managed through GitHub repositories and Dropbox. The speaker avoids sharing personal Dropbox with OpenClaw due to privacy concerns, opting instead to set up a separate Dropbox account for OpenClaw.

Cost Considerations in Using Agents

• The speaker warns about potential high costs associated with token usage when interacting with agents, noting they exceeded $200 within two days of setup.

• They mention stories of accounts being shut down due to possible violations of Anthropic's terms of service, leading them to maintain separate API tokens for personal use and agent operations.

Optimizing Agent Performance

• The speaker uses a Cloud Max plan for personal tasks while running OpenClaw agents on separate API tokens through Open Router, which helps optimize model selection based on task requirements.

• Emphasizes the importance of optimizing which models are used for specific tasks, balancing cost against performance needs.

Chat Tools and Multi-Agent Configuration

Initial Chat Tool Experience

• Initially using Telegram for agent communication proved uncomfortable due to markdown formatting issues; thus, they transitioned to Slack which offers better support for threaded replies and markdown content.

Building a Team of Agents

• The speaker created a multi-agent configuration consisting of four distinct agents: Claw (system admin), Bernard (developer), Vale (marketing), and Gumbo (general assistant), each operating as individual Slackbots.

• Each agent is assigned specific models based on their roles—Opus for reasoning-intensive tasks and Sonnet for speed-oriented tasks—allowing delegation among sub-agents when necessary.

Managing Tasks and Custom Solutions

Centralized Task Management

• All agents share one workspace that allows access to shared memory, facilitating management configurations across all agents from one location.

• An identity.md file is utilized creatively to define multiple identities corresponding to each agent's unique personality traits inspired by characters from the band Gorillas.

Challenges Leading to Custom Development

• Difficulties with OpenClaw’s built-in cron system prompted the creation of a custom dashboard that consolidates scheduled tasks per agent while tracking token usage effectively.

• A simple Rails app was developed in about a day, providing an interface that connects with OpenClaws Gateway for comprehensive management oversight.

Building Tools for Enhanced Productivity

Introduction to Agent Utilization

• The speaker is developing an app for editing and reading markdown files, aimed at managing agent access efficiently.

• A key question arises: what specific functions will the agents serve in the business?

Identified Areas for Agent Application

Content Management

• The speaker emphasizes that they only publish content when there is something meaningful to share, but much valuable work goes unshared.

• Systems are being built to allow agents to observe and capture more project-related activities for broader sharing.

Development Tasks

• The speaker enjoys hands-on development but has assigned their developer agent, Bernard, to handle backlog issues and production errors during busy times.

Administrative Work Automation

• Project management tasks such as scheduling and documentation are seen as bottlenecks; these should be automated or delegated.

• Gumbo, the general assistant agent, is tasked with alleviating these administrative burdens.

Insights from Reporting Use Cases

• The most exciting use case involves having agents identify trends and insights that can reveal blind spots in the speaker's work.

• This insight aids in teaching builders effectively and creating tools that address real problems.

Future Developments with OpenClaw

• The speaker acknowledges OpenClaw's potential despite its early-stage development, highlighting its significance among AI tool builders.

• Emphasizes the importance of being an early adopter of such systems as they become increasingly prevalent.

Essential Skills for Builders in 2026

• Builders must explore new tools actively to thrive amidst rapid changes; this adaptability is crucial for success.

• The speaker mentions five essential skills needed to transition from feeling overwhelmed by change to thriving within it.