The Orchestration Is the Product (And Now We Have Proof)

Dave Eichler — April 8, 2026 — AI, orchestration, governance, Claude Code, Conway, Karpathy, PRISM-IR, #NOKINGS

Why Anthropic's leaks confirm what some of us have been building toward for months — and why the window to act is closing.

On March 31, 2026, Anthropic accidentally shipped 512,000 lines of source code to the public npm registry. Not model weights. Not training data. The orchestration layer.

Most people assumed Claude Code was a polished CLI wrapper around the Anthropic API. The leak revealed something else entirely: a full multi-agent production system with circuit breakers, retry logic, graceful shutdown handling, sub-agent spawning, task queues, memory consolidation, and a three-layer architecture for managing "context entropy" — the tendency for AI agents to become confused or hallucinatory as long-running sessions grow in complexity.

512,000 lines of whack-a-mole.

The industry finally saw what I've known for months: the model is commodity. The orchestration is the product.

Then, a week later, Nate B Jones dropped his analysis of Anthropic's next leaked product — Conway. And suddenly the full picture snapped into focus.

The Conway Threat

Conway is the always-on agent. It watches you work. It learns which emails require your judgment and which can be auto-responded. It knows which Slack threads matter. It preps your board meetings before you've typed a word. It accumulates behavioral context over months until it understands how you think.

Nate identifies the threat correctly: behavioral lock-in.

Not your files. Not your Slack messages. The patterns the agent learned by watching you. The model of how you work. There's no CSV export for that. No migration consultant. When you leave Conway after six months, you leave your brain behind.

His call to action:

"The policies around behavioral context portability should ship before Conway does, not after."

He's right. But policy isn't enough.

What the Leaks Prove

The Claude Code leak showed us exactly how Anthropic solved the orchestration problem. Three-layer memory architecture. MEMORY.md as a lightweight index of pointers. Topic files fetched on-demand. Raw transcripts never fully loaded — just grep'd for specific identifiers. "Strict Write Discipline" where the agent updates its index only after successful file writes. "Self-Healing Memory" that treats its own knowledge as hints requiring verification against the actual codebase.

And underneath all of it: KAIROS — an autonomous daemon mode where the agent runs background tasks, performs "memory consolidation" while you're idle, merges observations, removes contradictions, and converts vague insights into facts.

This is Conway's ancestor. The always-on agent that learns how you work is really an orchestration system that accumulates patterns. 512,000 lines of code playing constant whack-a-mole with model chaos.

The leaks prove three things:

First, generative AI can remain creative — even hallucinatory. That's a feature, not a bug. The process built to harness it must never be idempotent machinery. It must be governed learning. Self-improving systems that get better over time, not brittle pipelines that break when the model drifts. Anthropic knows this. That's why they built KAIROS with "autoDream" and memory consolidation.

Second, the orchestration layer is where the real value lives. As one analyst put it after the leak: "What reached the public was not the model itself but the full orchestration layer around it. That layer — the workflow logic, memory management, tooling integrations and agent coordination — is where a foundation model becomes operationally viable."

Third, Anthropic doesn't want you to see how any of this works. They published MCP as an "open standard" while the real coordination logic stays proprietary. Conway will use MCP, but Conway's CNW.zip extension format sits on top — proprietary tools that only work inside Anthropic's walled garden. The open layer is plumbing. The proprietary layer is value.

It's the Google Play Services pattern. Android is open. The valuable stuff isn't.

The Architecture I Built Instead

I started building a different system months ago. Not because I predicted the leaks — but because I intuitively understood that whoever owns the orchestration owns the user.

So I built the orchestration on MY side.

File-based protocols. Every task, every response, every decision flows through files that live on my machine. The LLM is a stateless worker. It reads a task file, does work, writes a response file. The context, the patterns, the accumulated intelligence — that lives in my file system, not Anthropic's servers.

Chat transcripts I possess. When I talk to Claude in the browser, I download the transcript. When I talk to Claude through my own interface, those logs hit MY database. When I work with Claude Code, those sessions save to MY machine. The conversation history isn't trapped in their infrastructure. I own it. I possess it. I can grep it, analyze it, train on it, migrate it.

Orchestration I control. The prompt injections that fix Claude's behavioral issues? I write them. The shims that prevent scope creep? Mine. The retry logic, the validation gates, the escalation paths? All in my codebase, versioned in my repo, running on my infrastructure.

When Claude misbehaves — and it does, constantly — I don't file a support ticket. I adjust my orchestration. I update my injection. I route around the damage.

The model is a utility. The intelligence is mine.

The Convergence

The timing isn't coincidence. Everyone's waking up at the same moment.

March 31: Claude Code leaks. 512,000 lines of orchestration exposed.

April 3: Andrej Karpathy — co-founder of OpenAI, former Tesla AI Director — drops his "LLM Wiki" pattern. His core insight: "A large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge."

Karpathy's approach rejects RAG entirely. Instead of retrieving chunks at query time, the LLM compiles your documents into a persistent, interlinked markdown wiki. It reads your sources, extracts concepts, creates backlinks, flags contradictions, and maintains the whole thing automatically. The wiki is the knowledge base. The LLM is the librarian. Nothing is re-derived on every query — the synthesis is already done.

His framing is perfect: "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase."

He even runs "linting" passes — health checks where the LLM scans for inconsistencies, missing data, orphan pages. That's governance. That's exactly what my GateEnforcer does, just different vocabulary.

And then this response in his thread, from entrepreneur Vamshi Reddy:

"Every business has a raw/ directory. Nobody's ever compiled it. That's the product."

Karpathy agreed. He called his setup "a hacky collection of scripts" and said there's room for "an incredible new product."

I'm building that product.

The Missing Layer: Collective Intelligence

Tiago Forte gave us the philosophy — "Building a Second Brain," the CODE framework (Capture, Organize, Distil, Express) that pioneered personal knowledge management.

Nate B Jones gave us the AI-native implementation — Open Brain, a single database any AI can read from and write to. Your thoughts. Your context. Your captured knowledge.

Andrej Karpathy gave us the architecture — the LLM Wiki pattern where the model compiles and maintains your knowledge base, not just retrieves from it.

All three are necessary. None is sufficient.

Forte's system is your philosophy. Nate's Open Brain is your memory. Karpathy's wiki is your knowledge base. But what about the collective intelligence? What about the patterns we ALL discover? The orchestration tricks that work? The failure modes that kill you? The shims that fix Claude's weird JSON handling or GPT's tendency to hallucinate tool calls?

Anthropic just accidentally published 512,000 lines of this knowledge. It's being analyzed, forked, rewritten in other languages, studied by every competitor on the planet. The orchestration secrets are out — but only because of an accident.

What if we published them on purpose?

The Global Commons

Everything is a wiki. And we all contribute.

The DEIA Federalist Papers — 34 documents on human-AI governance, co-authored with Claude and GPT under the pseudonym PUBLIUS — are the philosophical foundation. They establish principles like:

But philosophy doesn't ship. Patterns do.

The Global Commons is where the community shares what actually works:

This is the knowledge Anthropic was hiding behind their proprietary layer — until the leak. The stuff that makes persistent agents actually work. The trial-and-error, the failure modes, the whack-a-mole solutions.

Karpathy runs linting passes on his personal wiki to find contradictions and inconsistencies. The Global Commons runs the same pattern at scale — the community identifies what works, what fails, what's been superseded. The knowledge compounds.

We publish it in the open. On purpose.

Not because we're altruistic. Because the community can govern these models better than the vendors can. We're the ones using them in production. We're the ones discovering the edge cases. We're the ones building the workarounds.

Anthropic has a few hundred engineers. The community has millions of practitioners. Who do you think will figure out better orchestration patterns?

The Translation Layer

There's one more piece: PRISM-IR.

PRISM-IR is a process representation language — human-readable but precise enough that LLMs interpret it consistently. When you write governance patterns in PRISM-IR, they work the same way on Claude, GPT, Gemini, or whatever model ships next quarter.

This is the Rosetta Stone. Without a common language, we're all writing English prompts that work slightly differently on every model. With PRISM-IR, we're writing specifications that translate cleanly everywhere.

Karpathy's wiki uses markdown. The Global Commons uses PRISM-IR. Same principle — structured, interlinked, machine-readable — but designed for governance patterns instead of personal knowledge.

The commons becomes actually useful when the patterns are portable.

The Four Layers

So here's the architecture that defeats Conway's lock-in:

Layer 1: Possession — Your data lives on your infrastructure. The LLM vendors get stateless API calls. They never accumulate your context because your context never lives on their servers.

Layer 2: Orchestration — The coordination logic runs on your side. When Claude Code ships a new version that breaks your workflow, you adjust your orchestration and keep moving.

Layer 3: Commons — The community publishes governance patterns in the open. Not as proprietary extensions (CNW.zip). In public repos, with open licenses, accessible to anyone.

Layer 4: Translation — PRISM-IR makes patterns portable across models. The governance instructions don't get lost in translation because they're not trapped in natural language ambiguity.

Why Now

The Claude Code leak happened March 31. Karpathy published his pattern April 3. Conway is coming. OpenAI and Google are building their own persistent agent layers.

We have maybe 12-18 months before this calcifies into three proprietary silos with token interoperability theater and real lock-in underneath.

This is the App Store moment. In 2008, you could still build for the open web. By 2012, the center of gravity had shifted permanently to native apps in platform-controlled stores. The open web still exists. It lost.

The Federalist Papers are published: github.com/deiasolutions/federalist-papers-ai

PRISM-IR is published: github.com/deiasolutions/prism-ir

The architecture is being built: simdecisions.com

The commons needs contributors.

The Call

Nate asks which persistent agent layer you'll choose. Anthropic? OpenAI? Google?

That's a game of pick your jailer.

I'm offering a fourth option: yours.

Own your data. Possess your transcripts. Run your orchestration. Publish your patterns to the commons. Write them in a language that works everywhere.

The vendors want you to believe this is too hard. That you need their infrastructure, their Conway, their 512,000 lines of proprietary whack-a-mole.

You don't.

The LLMs are stateless workers. The intelligence is in the patterns. The patterns can be shared.

Build the commons before they brick off the path.

Dave Eichler is the founder of DEIA Solutions and the creator of SimDecisions. He's been building file-based AI orchestration systems since 2024, and the Claude Code leak confirmed he was on the right track. Find him on LinkedIn at @daaaave-atx.

#NOKINGS — Human sovereignty is non-negotiable.