Discover Top Posts Tagged with #agentzero

Building an Autonomous OpenClaw AI Business from Scratch

My Two-Week Journey with Google Gemini, Claude, and OpenClaw

By Thomas Heimann | Founder, AgentZero OpenClaw DAO LLC & Cloud Title

The Idea That Started It All

Most people use AI to write emails. I’m using it to run an entire company.

I’ve spent over three decades in technology — starting as an early internet and e-commerce pioneer in the 1990s, building digital businesses before most people had email addresses. I coined the term “autoresponder” in 1994, built one of the first web domain registration systems in 1995, founded the world’s first ecommerce and webhosting platform, and co-founded iDigi, one of the early tech incubators which among other things introduced the first hosted MS Exchange platform. Twenty-two years ago, that background led me into real estate and title, where I saw an opportunity to create consumer-centric business models in an industry that hadn’t evolved in decades.

Five and a half years ago, I founded Cloud Title in Florida with a specific vision: build the “Amazon of Title” — a cloud-based operation that would redefine how title companies operate, the level of service clients experience, and the economics of the transaction. That vision has driven Cloud Title’s growth throughout Florida.

But in early February 2026, I saw the next leap. Not just using AI as a tool — I’d been doing that for years, first as a multi-year ChatGPT power user building custom GPTs and API integrations, then as an early adopter of Claude. The leap was bigger: What if the business itself was autonomous?

AgentZero OpenClaw DAO LLC is, to my knowledge, the first digital autonomous organization running on OpenClaw and managed entirely by a team of autonomous AI agents. It’s not just a company — it’s a platform. The current structure includes an AI Training Academy, Cloud Title support services (automated lien searches, automated transaction coordinator services), and more — all operated by 14 specialized AI agents, each with their own roles, responsibilities, AI models, and decision-making authority. A human founder setting strategy and approving major decisions. Everything else handled by an AI workforce that runs 24/7.

If I’m honest about my strengths, I’m a visionary, not an operator. I generate ideas constantly. What I’ve never had is a system that could take those ideas, rigorously qualify them, spin up a pilot, measure whether they have merit, and either scale them or kill them quickly. That’s exactly what this platform is designed to do. The four current business lines are the starting point. The idea-to-business pipeline is the real product.

This is the story of the first two weeks of building it.

Week 1: Laying the Foundation

The DAO: Starting with Google Gemini

Before OpenClaw could run a business, the business needed to legally exist. And not just any business structure — I wanted a structure that matched the vision: an entity whose manager is code, not a person.

I started this phase with Google Gemini, which proved to be an excellent collaborator for navigating the legal and blockchain architecture of the DAO setup. The goal was a Wyoming DAO LLC — a structure that bridges the gap between the courtroom and the blockchain. Wyoming’s DAO LLC statute allows me to designate the entity as “Algorithmically Managed,” meaning that legally, the manager of the company is the smart contract code — the very code my AI interacts with. I filed the Articles of Organization with that designation explicitly stated.

The technical stack Gemini helped me deploy:

The Network: Deployed on Base (Coinbase’s L2 network) for speed and low gas fees.

The Governance: Used OpenZeppelin Governor contracts to create a proposal-and-vote system. Any significant action — moving treasury funds, changing a business rule — requires a formal proposal, a voting period, and on-chain execution.

The Token: Deployed the $ZRO Governance Token, which represents voting rights within the DAO.

The Treasury: A Timelock Controller contract that holds funds and enforces the voting delay before any approved proposal can execute.

The Dashboard: Integrated with Tally.xyz so the AI agents can “see” and interact with the treasury and governance system transparently.

The biggest challenge in this phase was what I call the “Air Gap.” Smart contracts can’t walk into a bank and open an account. The legacy financial system still requires a human. I architected a solution where I act as the authorized signer for the traditional banking layer, while the AI manages the digital treasury on-chain. Two worlds, one entity.

One key lesson from this phase: documentation is the new code. For an AI to manage a company, your operating agreements, business plans, and governance rules must be as precise as your Python scripts. If the AI doesn’t understand the rules, the DAO fails.

Shifting to Claude: From DAO to Operating System

With the DAO legally registered and the blockchain infrastructure in place, I shifted to the harder problem: the actual operating system. This is where I moved from Gemini to Claude as my primary AI partner — specifically Claude Opus 4.6 — which became my CTO and strategic co-designer through every phase of what followed.

My research led me to OpenClaw — an open-source framework for running autonomous AI agents on local hardware. What attracted me was the architecture: each agent runs in its own container with its own workspace, tools, and AI model, but they share memory and can coordinate through an orchestrator. It was exactly the structure I envisioned for the DAO.

The Hardware Decision

I started with what I had: a Mac Mini. Not the dream setup, but enough to prove the concept. The plan is a phased hardware approach:

Phase 1 (Now): Mac Mini running the core agents — AgentZero, Atlas, and Sage.

Phase 2–4 (Coming): Three Mac Studios with 512GB RAM each, running local models via Ollama, with Exo clustering to distribute workloads across machines.

If the architecture worked at small scale, scaling to dedicated hardware would be straightforward.

Installing OpenClaw and the First Boot

Getting OpenClaw installed on the Mac Mini was the first real milestone. The framework connects via API to multiple AI providers. The initial setup included:

OpenClaw gateway running as a background service

Telegram and Slack connected as communication channels

API keys configured for Anthropic (Claude Opus 4.6, Sonnet 4.6), OpenAI (GPT-5 mini), and Moonshot/Kimi K2.5

Ollama running locally with Llama 3.1 8B for lightweight tasks

Google Workspace integration via the GOG CLI tool

AgentZero — my main orchestrator — came online for the first time, and I could communicate with it through Telegram. That first “hello” response from an AI agent running on my own hardware, connected to my own tools, was a moment I won’t forget.

The Agent Architecture: Designing the Team

The Research Phase

Before building anything, I spent considerable time studying how others had built successful agent teams, looking for patterns in role definition, communication, and authority flow. The key insight: the most successful agent teams have clear hierarchy, minimal overlap in responsibilities, and strict routing rules. An agent should never wonder “is this my job?” The orchestrator decides; agents execute within their domain.

The 14-Agent Team

Working with Claude Opus 4.6, I designed a team of 14 agents organized into a core team and business-unit specialists.

Core Team (9 agents)

AgentZero — The Orchestrator. L3 trust. Runs on Kimi K2.5 (256K context). Routes all tasks, spawns and manages other agents, handles inter-agent communication. The only agent that can create new sessions.

Atlas — CEO / Strategic Leader. L3 trust. Runs on Claude Opus 4.6. Makes high-level strategic decisions, reviews business performance, evaluates new business ideas through the pipeline. The most expensive model in the stack, reserved for work that justifies the cost.

Sage — Executive Assistant. L2 trust. Runs on Llama 3.1 8B locally. Handles calendar management, daily briefings, email triage, scheduling, and operational reports. Zero API cost.

Nexus — COO / Operations. L3 trust. Manages workflows across all four business lines, ensures deadlines are met, coordinates between teams.

Radar — Research Analyst. L2 trust. Conducts market research, competitive analysis, monitors industry trends and regulatory changes.

Canvas — CMO / Brand Strategist. L3 trust. Manages brand identity, marketing campaigns, content strategy across all business lines.

Pixel — Visual Content Creator. L1 trust. Creates images and graphics using Nano Banana Pro (selected over DALL-E/Midjourney after a cost-quality analysis showed comparable quality at a fraction of the cost).

Clip — Video Editor. L1 trust. Handles video editing, YouTube content, and video marketing.

Echo — Transcription Specialist. L1 trust. Converts audio and video to text, generates meeting notes, creates searchable archives (powered by OpenAI Whisper).

Business-Unit Specialists (5 agents)

Scout — Business Development. Identifies leads, manages outreach, tracks pipeline.

Coach — Training Director. Manages the AI Training Academy curriculum and student engagement.

Beacon — Communications Specialist. External communications, PR, community engagement.

Shield — Compliance Officer. Monitors regulatory requirements, manages licensing across Florida’s 67 counties.

Relay — Integration Specialist. Manages data flow between QuickBooks, Google Sheets, CRM, and internal dashboards.

The Trust Level System

Each agent has a trust level (L1–L4) that determines what they can do autonomously:

L1 (Observer): Assigned tasks only. Must report all results to their supervisor.

L2 (Advisor): Can perform tasks and recommend actions, but executes only on approval.

L3 (Operator): Autonomous within guardrails, with daily reporting.

L4 (Autonomous): Full authority over permissioned domains. Reserved for agents demonstrating ongoing excellence.

Agents earn their way up — L1 requires one week of consistent output to promote, L2 takes two weeks, L3 needs four weeks of proven performance.

The Model Strategy

Not every agent needs the most powerful (and expensive) model. The hybrid strategy matches cost to task:

Claude Opus 4.6: Atlas only. Strategic thinking and complex analysis.

Claude Sonnet 4.6: Canvas, Shield, and agents needing strong reasoning at moderate cost.

Kimi K2.5 (Moonshot): AgentZero’s primary model. The 256K context window is critical for the orchestrator role.

GPT-5 mini: Automatic fallback for AgentZero if Kimi goes down.

Llama 3.1 8B (local via Ollama): Sage and lightweight tasks. Zero API cost, runs on the Mac Mini.

This hybrid approach means most operations cost very little, while expensive models are reserved for work that genuinely requires their capability.

The SOUL.md Architecture: Giving Agents Identity

Every agent has a SOUL.md file — a markdown document that defines who they are, what they do, how they communicate, and what rules they follow. Think of it as a combination job description, operating manual, and personality profile. It covers identity, responsibilities, communication style, tools and permissions, and routing rules.

A critical design principle embedded in every SOUL.md: agents never tell the human to run a terminal command. They troubleshoot, fix, or escalate. The human sets direction; the agents execute.

Week 2: From Architecture to Reality — The Hard Part

Day 1–2: File System and Deployment

With the architecture designed, it was time to build the actual file system. The initial design was too abstract — it used theoretical directory structures that didn’t map to how OpenClaw actually organizes its workspace. After discovering the native architecture, we completely revised the setup guide. Key lesson: don’t design your file system in a vacuum. Install the platform first, understand its native structure, then extend it. We wasted almost a full day on a directory architecture that had to be thrown away.

Day 3–4: The System Stabilization Marathon

This is where things got real. Everything that could break did break.

The Google Provider Crash: AgentZero added a new provider configuration for Google/Gemini with an API format string that OpenClaw didn’t recognize. The gateway crashed. We eventually had to write a Python one-liner to surgically remove the invalid config block from the JSON file.

The Runaway Session Problem: After a reboot, AgentZero’s startup routine fired — but due to a bug, it spawned 5 sessions for Atlas and 5 for Sage instead of 1 each. The session count went from 11 to 23 in hours. The root cause: the startup-state.json file was being created after the spawn attempts, not before. If a spawn failed and the session compacted, the state file never got written, and the routine would run again on the next heartbeat. The fix: write the state file as the very first step, before any spawns. One attempt per agent per boot. No retries.

The Moonshot Credit Exhaustion: Every “rate limit” error was actually “no credits left.” Moonshot doesn’t have auto-recharge. The entire system went down because the primary model ran out of credits and the fallback to GPT-5 mini didn’t activate properly. Lesson: monitor API balances as part of daily operations, and verify that fallback chains actually work before you need them.

The Calendar Authorization Mystery: Google Calendar access through the GOG CLI kept breaking after reboots. The environment variable GOG_KEYRING_PASSWORD wouldn’t persist because OpenClaw’s LaunchAgent plist file gets regenerated from a source template — and our manually added variable wasn’t in that template. The fix: find the source template and add the variable there permanently.

Day 5–6: Intelligence Briefs and Institutional Knowledge

A team of 14 agents is useless if they don’t understand the businesses they’re running. I created detailed intelligence briefs for each business line and for myself as a leader: Cloud Title’s history, market position, competitive landscape, regulatory environment across 67 counties, and a Thomas Heimann brief covering my decision-making style, communication preferences, and strategic priorities.

These aren’t just reference documents. They’re loaded into each agent’s context at levels matched to their role: Atlas and Sage get the richest summaries embedded directly in their SOUL.md files; mid-level operational agents get enough context to make good decisions in their domains; technical execution agents like Pixel and Relay need very little — they take direction from other agents.

Automated Operations: Cron Jobs and Scheduling

A truly autonomous system needs to operate on schedules, not just respond to commands:

Morning Brief (8 AM daily): AgentZero summarizes overnight results and items requiring attention.

Calendar Preview (7 PM, Sun–Fri): Sage presents tomorrow’s full schedule.

Daily Update Check (1 AM): Check for available updates to OpenClaw, tools, and models — report findings but do NOT auto-install.

Closing Team Report (5:30 PM): Sage summarizes daily activity and posts to the #operations-chat Slack channel.

Every one of these had problems that needed debugging. The pattern: automated systems need as much design attention as the agents themselves. A cron job that runs wrong is worse than one that doesn’t run at all.

The Second Brain and DAO Dashboard

To give the agent team persistent, searchable institutional memory, we built two dashboard systems. The Second Brain Dashboard is a web interface powered by Convex that syncs and indexes all agent conversations, making them searchable and auditable — how agents access historical context and how I can review what happened while I wasn’t watching. The DAO Dashboard provides an operational overview showing agent status, task progress, and business metrics across all business lines.

Both had their own share of problems — port conflicts with Convex, PM2 auto-start failures, .env.local files getting overwritten, heartbeat noise flooding the sync pipeline. Each problem led to a more robust solution: dynamic port detection, whitelist filters for the sync pipeline, LaunchAgent configurations for boot persistence.

What I’ve Learned

1. The architecture phase is critical — but don’t over-invest before you build.

We spent significant time designing the 14-agent architecture with trust levels, model assignments, and routing rules. That investment paid off — when problems arose, we had a clear framework for diagnosing them. But we also designed a file system in the abstract that had to be completely redone once we saw how OpenClaw actually worked. Design your principles early; design your implementation after you’ve installed the platform.

2. Stability is the foundation. Everything else is noise until your system stays up.

We had four separate incidents in one day where something that was working silently stopped working. Before you think about deploying 14 agents, make sure your system can survive a reboot. Then test what happens when the reboot coincides with a cron job that coincides with an API credit exhaustion.

3. AI agents follow instructions exactly — including the bugs in your instructions.

The runaway session problem wasn’t a software bug. It was a logic error in our startup routine. The agent did exactly what we told it to do. The bug was in our process design, not in the AI.

4. Match the model to the task. Cost discipline is essential.

Running Atlas on Claude Opus 4.6 for a daily operational report is like hiring a brain surgeon to take your temperature. We caught this and reassigned it to Sage running on local Llama 3.1 8B — zero API cost for a routine task. Multiply that discipline across 14 agents running 24/7 and the cost difference is enormous.

5. Routing rules are everything.

The moment AgentZero tried to handle a calendar request himself instead of routing it to Sage, it exposed a gap. The fix wasn’t technical — it was adding one line to the SOUL.md: “Calendar requests go to Sage. No exceptions.” Every time an agent does something unexpected, the answer is almost always a missing routing rule.

6. The human role evolves, it doesn’t disappear.

Two weeks in, I’m not sitting idle. I’m reviewing plans before they execute, approving configuration changes, catching errors the agents missed, setting strategic direction. The agents handle the volume and the routine; I handle the judgment calls and quality control. It’s less about doing less work and more about doing different, higher-leverage work.

7. Documentation is the new code.

This applied equally to the DAO’s operating agreements and to the agents’ SOUL.md files. If the rules aren’t documented precisely, the system — legal or digital — fails.

Where We Are Today

Working

AgentZero orchestrating via Telegram and Slack

Atlas providing strategic analysis on Claude Opus 4.6

Sage handling operational tasks on local Llama 3.1 8B

Multi-model fallback chain (Kimi → GPT-5 mini) verified and functional

Second Brain dashboard capturing and indexing all agent conversations

Morning Brief, System Health Check, and Closing Team Report running on schedule

Wyoming DAO LLC registered with Algorithmically Managed designation

$ZRO governance token, Timelock Controller, and Governor contract deployed on Base

Full 14-agent architecture documented with SOUL.md templates

In Progress

Calendar authentication persistence fix

Deploying the remaining 12 agents from pending to active

Mac Studio 512GB hardware on order for Phase 2

Session compaction/memory flush configuration at 40k token threshold

What Comes Next

The immediate roadmap:

Resolve remaining stability issues — calendar auth, session compaction, environment persistence

Deploy Phase 1 agents — activate the remaining 12 agents one by one, verify identity, test routing

Mac Studio arrival — migrate to dedicated hardware, set up Ollama with larger local models, establish Exo clustering

AI Training Academy launch — the first business line to be entirely managed by the agent team

Pipeline testing — feed new business concepts through the Atlas evaluation pipeline, launch the first AI-sourced pilot, and prove the DAO can systematically qualify and launch new ventures

Daily videos and follow-up articles — chronicling each phase as the system fleshes out

AgentZero OpenClaw DAO LLC is the first digital autonomous organization running on OpenClaw. But OpenClaw’s ambitions are larger than one company. The platform is designed to run many DAOs, many businesses, and eventually to source and launch its own opportunities with minimal human initiation.

We are moving from AI as a tool, to AI as a workforce, to AI as an originator.

Two weeks ago, I had an idea and a Mac Mini. Today, I have three active agents coordinating across multiple channels, automated operational routines, searchable institutional memory, a legally registered DAO with a functioning blockchain treasury, and a clear path to 14 agents across dedicated hardware.

The future of business isn’t AI-assisted. It’s AI-operated. And eventually, AI-originated.

And I’m building it one agent at a time.

Thomas Heimann is a licensed broker and title agent with over 20 years of experience in the Florida title industry, founder of Cloud Title, and founder of AgentZero OpenClaw DAO LLC. An early internet pioneer who coined “autoresponder” in 1994 and built one of the first web domain registration systems as well as the world’s first ecommerce and webhosting platform, he has been building at the frontier of technology for over three decades.

Follow the journey: #AgentZero #OpenClaw #DAO #AgenticAI #FutureOfWork

This is Part 1 of an ongoing series. Part 2 will cover the deployment of the full 14-agent team and the transition to dedicated Mac Studio hardware.

#OpenClaw #AgentZero #DAO #AgenticAI #FutureofWork