The Matrix Manifesto
The age of prompt-and-pray is over.
A statement on why reliable AI agency requires new infrastructure — not better prompts.
I. The problem isn’t the model.
Modern language models are remarkable. They can synthesize knowledge, write code, and reason across domains in ways that would have seemed impossible five years ago. But they were designed as text completion systems — not as reliable agents that take action in the world.
When we ask these systems to take irreversible actions — deploying code, executing transactions, mutating databases — we are relying on probabilistic text generation to produce deterministic correctness. This is the original sin of the current agent paradigm. And it’s fixable.
II. Intent is the missing abstraction.
Between natural language and system execution, there is a gap. That gap is where hallucinations live. The solution is not a better model — it’s a new layer: a typed, inspectable, correctable representation of what the agent intends to do before it does it.
We call this the Intent IR. It is the core abstraction of Matrix. Every action that flows through our system is first materialized as a typed intent object — with a closed verb vocabulary, structured parameters, and a declared reversibility class — before any side effect occurs.
III. Two rails, not one.
Not all agent work is equal. Writing a test, summarizing a document, or browsing the web is fundamentally different from deploying a smart contract or executing a financial transaction. We refuse to treat them the same.
Neo handles the reversible world: fast, permissive, memory-augmented, and deeply capable. MCL handles the irreversible world: typed, deterministic, and auditable to the byte. The handoff between them is automatic. The separation is absolute.
IV. Memory is not optional.
The stateless request-response model of LLM APIs is a consequence of how these systems were built, not a design choice. Agents that operate in the world need persistent, typed, cryptographically verifiable memory. They need to know what they’ve done, what’s changed, and what they’ve been trusted with.
Cortex — Matrix’s per-actor memory graph — is our answer. Merkle-anchored snapshots mean every state transition is verifiable. Append-only journals mean no action is ever silently lost.
“The agents that survive the next decade will not be the most capable. They will be the most trustworthy. Matrix is the infrastructure for that future.”
— Matrix founding team