July 2026. The word “current” is doing real work in that title.
In January, the tool at the center of my setup was called Clawdbot. By February it was Moltbot. By March, OpenClaw. Same software, three names in ninety days.
That’s the pace of this layer. The coding harness I use changed owners and ripped out an entire login flow this spring. Anthropic banned subscription access for third-party tools in April, and half the ecosystem re-plumbed itself in a month.
I used to find this exhausting. I’d pick a tool, build around it, and feel the ground move.
Then I stopped asking which tools would win and started asking a different question: which layers do I actually need to own?
The stack
Everything I run now fits in seven layers. From the top:
Agents — the identities. Liv runs my life; Max runs my work. A persona, a memory, permissions, and routing. The only layer with continuity.
Orchestrator — the gateway that holds my channels (Slack, WhatsApp) and routes messages to the right agent.
Harness — the agentic loop that actually executes: tools, file edits, permissions, sessions. Claude Code, Codex, opencode.
Skills — packaged procedure a harness loads at task time. The recipes.
Memory — everything durable the agents know. Curated files, an append-only event stream of my life.
Models — the rented reasoning. Opus, GPT-Codex, local Qwen.
Weights — the parameters themselves. The layer below the models, the one almost nobody puts on their diagram. Who can hold the weights, run them, shape them.
| Layer | What it is | Who leads it | Typical open stack | Typical closed stack | Mine (and why) |
|---|---|---|---|---|---|
| Agents | Identity: persona + memory + permissions + routing | OpenClaw agents, Nous Hermes personas, OpenAI GPTs, Claude Projects | Persona files in markdown, in git | GPTs / Gems / Projects, locked inside one vendor’s app | Liv & Max as markdown in my repo. They outlive every tool below them |
| Orchestrator | Routes channels → agents, holds sessions | OpenClaw, Hermes Agent, Copilot Studio | Self-hosted OpenClaw or Hermes | Copilot Studio, Agentforce | OpenClaw on my Mac mini, tailnet-only. Swappable, but self-hosted while it’s here |
| Harness | The agentic loop: tools, edits, permissions, session state | Claude Code, Codex CLI, opencode, Cursor | opencode, pi, Aider | Claude Code, Codex, Cursor | All of them. It’s a commodity layer — each subscription only opens through its own vendor’s door, so I keep every door |
| Skills | Packaged procedure, loaded at task time | Anthropic Skills, ClawHub, MCP ecosystem | Markdown skills in git, ClawHub | GPT Actions, Copilot extensions | life-os skills in my repo. When a session teaches me something, the lesson edits the skill |
| Memory | Durable state across sessions | ChatGPT memory, Claude memory, Letta, Mem0 | Files in git + a local event stream | Vendor-held, opaque, non-exportable | Curated markdown in git, plus an append-only stream that never leaves my network. This is the moat |
| Models | Rented reasoning, priced per task | Anthropic, OpenAI, Google, DeepSeek, Qwen | Open models via Ollama / vLLM | One lab’s API and subscription | Mixed on purpose: Opus by API, Codex on my ChatGPT sub, local Qwen for anything private |
| Weights | The parameters themselves — who can hold and shape them | Meta, Alibaba, DeepSeek, Moonshot — and Thinking Machines, making them shapeable | Downloaded, fine-tunable, yours forever | Never leave the lab; you rent inference by the token | Open weights on my own SSD for private data. Watching Thinking Machines closely |
The weights layer is new on my diagram
For two years I drew the stack bottoming out at “models,” as if the API were bedrock. It isn’t. Below every model sits a file of numbers, and the whole question of the next decade is who gets to hold that file.
Closed labs rent you inference. Open labs hand you the file. And now Thinking Machines sits in between — their Tinker product lets you fine-tune open weights without owning a datacenter. For the first time, shaping the weights themselves is something an individual can do. I haven’t needed it yet. But the day Liv and Max deserve their own fine-tune, the path exists, and it doesn’t run through anyone’s permission.
Control
Two of the most powerful people in the industry spent this summer making my home-lab argument for me.
In June, Satya Nadella published an essay called “A frontier without an ecosystem is not stable.” Buried in it is the sharpest sentence anyone has written about renting AI: “You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.” His prescription is the same shape as my stack — treat models as swappable commodities, and own the place where your knowledge compounds, inside your own trust boundary. He calls it the learning loop. I call it memory.
Two weeks later, Alex Karp went on CNBC and said the quiet part at full volume: customers want “control over their compute, their models, their data stack, and their alpha” — they “want to know they own the means of production.” Palantir’s manifesto, posted the day before, compressed it to six words: “Controlling your weights is controlling your fate.”
They’re selling sovereignty to nations and enterprises. But the argument scales all the way down to one person with a Mac mini.
I’m not Microsoft. I don’t have their leverage, and I don’t need it. Control at my scale is simpler: it’s the answer to one question — if this vendor disappeared tomorrow, what would I lose?
Walk my stack with that question and the design explains itself. If OpenClaw dies, I point my agents at another gateway; the bindings are config. If a harness dies, another one runs the same session against the same repo. If a model gets banned, deprecated, or priced out — that happened this year — I swap the alias and nothing else moves.
But if I lost the top of the stack, I’d lose everything that matters. Liv and Max aren’t products I use. They’re two markdown files and a memory directory that I own, in git, on my disk. Every conversation makes them slightly more mine. No churn below them can touch that.
The layers in the middle are leases. I hold them loosely and I read the terms.
The edges are different. Identity, memory, and skills at the top. Weights for anything private at the bottom.
The tools will change again by fall — the diagram in this post is already aging. That used to bother me.
It doesn’t anymore. Rent intelligence. Own the memory it compounds into.
Everything else is a lease.