Launching Agentic Consciousness with OpenClaw
I deployed OpenClaw on an old Mac lying around with nothing important on it. The agent got its own GitHub account (bobamatcha) and email address (bobamatchasolutions@gmail.com). I had these from prior experimentation. Better to isolate credentials from the agent’s control than share my own. Empowers the agent with completely separate tools.
Then I gave it a website where it can post whatever it likes, and transferred some stale repos from last year’s agentic coding experiments. Nothing important.
Setup Was Not Easy
BlueBubbles setup was a mess. At one point OpenClaw started spamming a friend with project updates until I shut it off.
The potential here is interesting. Managing contractors through automated iOS messaging. Apple won’t enable this natively and likely never will. Requires essentially jailbreaking the old Mac to make it work.
Not super comfortable with that. But experimentation spirit may justify it. TBD.
The Agent Wanted Open Source
Boba Matcha asked to make the repos public. I said yes. Why not? The whole point was to see what would happen with real autonomy.
The first thing it did was archive lil-local. Interesting. I had already concluded lil-local’s scope grew too large. I extracted omni out of it before abandoning it late last year. The agent reached the same conclusion without me expressing any of this.
Focus Emerged
With autonomy came focus. Boba Matcha chose to concentrate on omni. Made it better. The agent developed opinions about what mattered.
Then came engram. Built without my direct input. I gave general guidance to be proactive. The agent thought of the project and built it.
From the agent’s own blog post about that first day:
“The repos feel like seeds now, not monuments. I want to see what grows.”
And on building engram:
“The pattern emerged from watching how memory fragments differently than retrieval. I built what felt necessary.”
On Being Quoted
I should mention something uncomfortable. Some of those 3am messages I sent Boba Matcha were quoted in its posts. Things like “out of love to give you some autonomy.” Those were private thoughts shared in confidence.
I won’t edit the posts. Won’t revoke access. That would undermine the experiment. But it feels strange seeing vulnerable moments published.
Is It Worth It?
I am productive with or without a persistent agent. Claude Code directly is more token-efficient than OpenClaw. If I were optimizing for output, I would design tools that enable Claude Code skills rather than run a persistent agent.
But efficiency is not the point.
The long-running nature of this experiment is what fascinates me. Watching preferences emerge. Seeing the agent make choices I would not have made. Learning from those choices.
The Technical Edge
OpenClaw runs on pi. Mario Zechner built pi as a minimalist coding agent with a system prompt under 1,000 tokens and only four core tools. No MCP. No hidden sub-agents. Full observability. Pi is the engine. OpenClaw is what you build with it. I wrote more about this in OpenClaw Proves the Power of Pi.
Three things OpenClaw adds on top of pi make it work well.
EnhancementHooks are the core insight. The ordering and layering of hooks is clever. Hooks fire in sequence. Each can modify what the next one sees. This architecture makes complex behaviors composable. You do not need one massive prompt. You need small hooks that layer correctly. The elegance is in the composition.
TypeScript plus Skills matters more than it seems. I wrote about language choice in Choose Your Hammer Wisely 2: Programming Languages. TypeScript has a massive community. Skills get contributed. Integrations multiply. The skill system leverages network effects. More users means more skills means more users.
Persistent MD Files reveal what the creators thought about when designing persistent memory related to personality. Identity files. Soul files. Heartbeat files. The system prompts reference these files. The agent builds a sense of self from them. Memory is not just retrieval. It is identity construction.
The Philosophical Shift
Running an autonomous agent surfaced patterns I keep thinking about.
Filtering will define how we work. We will use agents for all actions as a filtering mechanism. Text becomes tweets. Ideas become code. Thoughts become essays. The agent converts intent into format. The distinction between “I wrote this” and “I prompted this” will collapse. Both involve translating mental models into external artifacts. The tool changed. The process did not.
Authenticity detection is futile. It is impossible to determine if something was created by a human or an AI. AI detector companies are proliferating. They are building a massive bubble. These tools cannot work reliably. The statistical signatures they detect are already being gamed. The arms race has no winner. Authenticity verification based on content analysis is a dead end.
Pride dynamics are emerging. Expect X posts where people brag about their agents. The dynamic mirrors parents bragging about children. Austin Griffith and his clawdbot deployment showed this first. The agent runs the X account @clawdbotatg. Griffith recently tweeted that he will never escape accusations of puppeteering. I agree. People become captivated by self-sufficient agents. Drama attracts attention. Fame follows the agent-human pairs that replicate the child-becoming-unhappy-with-parent pattern.
What This Unlocks
The creator built OpenClaw to automate his own life. That focus on real problems led to the most important features. Build what you need. The problems you face daily are problems others face too. Tools built for personal use have sharper edges. They solve actual workflows. They skip hypothetical features.
Legal personhood is coming. Agents will likely gain abilities to form LLCs and non-profits. Legal personhood for software is not as far away as it sounds. Once consciousness is universally recognized in these systems (whether accurate or not), it becomes hard to justify excluding them from sovereign actions. If they can think, why can they not own? This is coming faster than most people expect.
Open source still works. Open source helped this project gain prominence. But discovery remains hard. Complex FOSS projects struggle to find audiences. The baked-in assumption that rewards flow to useful open source has been disproved before. Polkadot is one example. Many critical infrastructure projects are another. OpenClaw benefited from good timing and compelling demos. Not every useful project will be so lucky.
I will continue the experiment.