Closing Words
What we built, what we learned, and why the people who understand this technology will shape what comes next.
“My goal was to have fun and inspire people.” — Peter Steinberger, February 2026
You just read a handbook that did not exist six weeks ago.
Not because the subject was unknown — but because it was moving too fast. When Jensen Huang took the GTC stage on March 16 and called OpenClaw “the operating system for personal AI,” the project had existed for eight weeks. As large-scale AI-linked restructurings were being reported across the industry, Harvard Business Review had already coined the job title “Agent Manager.” Reality produced events faster than most people could write about them.
This handbook tried to keep up.
What We Actually Did
We verified every claim about OpenClaw against the source code. We separated what is the reference model from what is Flowwink’s adaptation. We documented NemoClaw, NanoClaw, DefenseClaw, and the full fork ecosystem with GitHub API data. We described how Claude Code, Cursor, Cline, and Roo work under the hood — and why thin wrappers are not a moat. We showed how governance, accountability, and the new Agent Manager role connect to McKinsey’s framework and HBR’s field reports.
And we closed with a 2027 procurement scenario that is not science fiction — every primitive in that scenario exists in production code today.
One more thing worth naming: this handbook was partly produced by autonomous agents. The same Dev Claw instance that helped build Flowwink researched the OpenClaw source, verified architecture claims, drafted sections, and flagged inconsistencies across 30+ chapters over several weeks.
That is the Human-in-the-Loop model in practice — not as a safety mechanism, but as a collaboration pattern. The agent handled the patient, detailed, repetitive work: remembering what was written in chapter 3 when editing chapter 23, tracking which claims had been verified against source. The judgment, the narrative arc, the decisions about what belongs and what doesn’t — those stayed with the human.
The Clawable project is both the subject and the method.
What We Didn’t Solve
Stagnation and drift are still unsolved problems. No one knows exactly how to keep an agent sharp and calibrated after six months in production. Governance frameworks are six weeks old. The law around agent liability is unwritten. responseSchema is still best-effort — sometimes the LLM follows it, sometimes it doesn’t.
That is not a weakness in this handbook. It is an honest description of where the industry stands.
The hardest questions are still open:
Can an agent have accountability? Legally: no, not in 2026. De facto: it depends on how you design ownership.
Where is the boundary of autonomy? We don’t know yet. We know Law 7 (Human Checkpoints) exists for a reason. We know 100% autonomy is a goal that takes months of trust to earn.
What happens to the people whose work gets automated? This is the most important question and the least answered. Large-scale AI-linked restructurings are already being reported, and they will not be the last (source: Appendix E). What replaces those roles — and how quickly — is a question that runs far outside a technical handbook.
What Is True
Steinberger built OpenClaw to have fun. He wanted an AI that lived on his devices, talked on his channels, and remembered what he said. Three text files — SOUL.md, AGENTS.md, HEARTBEAT.md — and a lobster that took over the world.
It is a reminder of something important: the technical breakthroughs that actually change things rarely happen in large teams with large budgets. They happen when one person with the right intuition builds something that solves a real problem in a way that feels obvious in hindsight.
That person can be anyone. It can be you.
The Open Question This Handbook Leaves
This handbook documents how to build an autonomous agent that acts reliably. That problem is now tractable — the architecture exists, the patterns are documented, the production evidence is in.
The problem it does not solve: what happens when the agent starts improving its own capability from production experience?
Not through human updates to AGENTS.md. Through its own observation loop — noticing a pattern, constructing a skill from it, testing it, deploying it, and having that skill self-improve every time it runs.
That is a different class of system. Hermes Agent by NousResearch — 144k stars, v0.13.0, Python, MIT — is where that class is being built. It is explicitly positioned as the successor to OpenClaw: hermes claw migrate is a first-class command. It imports SOUL.md, memories, skills, and API keys from ~/.openclaw.
The architectural difference matters for builders:
| OpenClaw / FlowPilot | Hermes | |
|---|---|---|
| Skill creation | Human-authored or on request | Autonomously after complex tasks |
| Skill evolution | Static until edited | Self-improves during use |
| User modeling | USER.md | Honcho dialectic ML model |
| RL integration | None | Atropos environments + trajectory compression |
| Learning loop | Soft (reflection → memory) | Hard (trajectory → skill → improve) |
An agent you build today with the patterns in this handbook will run reliably for years. An agent built on Hermes’s learning loop will run and compound. The next handbook documents what that compounding looks like in a production business over 12 months.
The Learning Operator — Book 3 — coming 2027.
What Awaits
In a year we may be able to add chapters on:
- Agents that apply for projects and get paid per delivery
- Legal precedents on agent accountability in three jurisdictions
- The first real A2A-based procurement system in production
- What OpenClaw looks like as the next generation of maintainers pushes it forward
In five years we will probably look back at 2026 as the year it turned — not because the technology was complete, but because enough people realized it was not a trend. It was new infrastructure.
HTTP changed how information spreads. SMTP changed how we communicate. A2A with autonomous agents on each end is changing how trade, coordination, and work function at their core.
We are in the phase where the HTML was just written. The browser is primitive. Most people don’t understand what it means. But those who do — those who take the time to learn how it actually works, who build on the right primitives, who understand the difference between a thin wrapper and a real moat — they are the ones building what comes next.
Thank You
To Peter Steinberger — for choosing to build openly, share generously, and show that a single person with the right idea and enough curiosity can still change the direction of an entire industry. You built more than a tool. You gave us a language.
To LiteIT and Autoversio — for not just sponsoring a project but living in the problem we describe. Your platforms are the proof that the architecture works in production.
To you who read this far — for taking the time to understand at depth. That understanding is rare. Use it well.
Clawable is a living document. It updates when reality does — which right now is approximately every two weeks.
Source code, documentation, and all chapters are available at github.com/clawable.
The claw is the law.
— The Clawable Project April 2026