A cooperative of brains: how agent orchestration and owned memory actually work

Why a single model is the wrong shape for real work

Most autonomous systems are built as one large model asked to do everything. That shape is convenient to demonstrate and difficult to trust. A single model has one context window, one set of instincts, and no natural place to draw a line between what it is allowed to decide and what it must escalate. When it makes a mistake, there is rarely a clean record of why, and almost never a way to reproduce the decision. For a business that has to answer to auditors, regulators and its own board, that is not a foundation. It is a liability wearing a demo.

We took a different route with Mickai. Mickai is a Sovereign Intelligence Operating System, a SIOS, and it is organised as a cooperative of specialist brains rather than one oracle. Each brain is good at a narrow thing. They work together under a governance layer that decides who acts, in what order, and what happens when two of them disagree. The result behaves less like a chatbot and more like a well run organisation, with roles, responsibilities and an audit trail that survives scrutiny.

Fifty brains, two families, one chain of command

The cooperative is made of 50 specialist brains. We split them into two families so that responsibility is always clear. Twenty five are domain brains, each carrying deep competence in a specific field of work. The other twenty five are operational brains, which handle the machinery of getting things done safely: routing, memory, verification, scheduling, and the housekeeping that keeps a long running system honest over weeks rather than minutes.

The important detail is not the count. It is the separation. Because a domain brain never also owns the rules that govern it, no single brain can quietly grant itself more authority. A finance brain can propose a payment. It cannot approve, sign and record that payment on its own. Those steps live in different hands by design, which is exactly how a careful human organisation splits duties to prevent quiet errors from becoming expensive ones.

• Domain brains carry the expertise and propose the work to be done.
• Operational brains route, verify, remember and schedule, keeping the system coherent over long horizons.
• A governance layer arbitrates between them and decides who is allowed to act.
• Every action ends up as a signed, replayable record rather than a vanished decision.

Deterministic arbitration, not a popularity contest

The interesting problems in multi agent systems begin the moment two capable agents reach different conclusions. Many systems resolve this by voting, by letting the loudest model win, or by asking yet another model to referee. Those approaches feel clever and behave unpredictably. The same inputs can produce different outcomes on different runs, which is intolerable when the decision moves money, changes a record, or touches a customer.

Mickai arbitrates deterministically. Given the same inputs, the same policies and the same state, the governance layer reaches the same decision every time. Precedence between brains is defined in advance rather than negotiated in the moment. When a proposal exceeds a brain's authority, or when two proposals conflict, the rules decide the outcome and the reasoning is recorded. This is what lets a customer treat the system as accountable software rather than a personality that has to be coaxed. Autonomy is real, but it runs inside a boundary that does not move because a model felt confident that day.

“Autonomy without governance is just an accident that has not happened yet. We made the governance the load bearing part.”

Memory the customer owns, layered by purpose

A cooperative of brains is only as good as what it remembers. Statelessness is fine for a single question and useless for real work, because real work spans days, involves earlier commitments, and depends on context that a fresh session cannot recover. So memory is a first class layer in Mickai, and it is organised by purpose rather than dumped into one undifferentiated pile.

There is short lived working memory for the task in front of the cooperative right now. There is durable memory that holds facts, decisions and commitments the organisation needs to keep. And there is the governed record of what happened and why, which is not memory the brains can rewrite to flatter themselves. Keeping these layers distinct matters. It means a brain can be given exactly the context it needs for a task without handing it the keys to everything, and it means the history of decisions stays trustworthy even as the working memory churns.

Above all, that memory belongs to the customer. It lives on the customer's own hardware, on premises and capable of running air gapped, with zero data egress and no public cloud round trip. Nothing about the cooperative's memory quietly leaves the building to be pooled, mined or repurposed elsewhere. The organisation that does the work is the organisation that owns the record of it.

Every action carries proof

Orchestration and memory only earn trust if you can check them after the fact. Every action a brain takes produces a cryptographically signed entry in the Open Audit Record. This is not a log file that anyone can edit and everyone quietly ignores. It is a tamper evident account of what was proposed, what was decided, which brain acted, and under whose authority. Because the signing uses post quantum cryptography (ML-DSA-65), those proofs are built to keep their integrity well into a future where today's cryptography has aged out.

This closes the loop that most agent systems leave open. When something goes right, you can show exactly how. When something goes wrong, you can trace it to the decision, the brain and the policy in force at the time, then reproduce it deterministically to understand and fix it. Governance, memory and audit stop being three separate features and start being one coherent story about accountable autonomy.

The moat under the architecture

This design is not only an engineering choice, it is protected work. We hold 104 filed UK patent applications carrying approximately 2,340 claims, with full specifications, claims and figures, building toward examination and grant. They cover the substance of how the cooperative is governed, how memory is layered and owned, and how the audit record is signed and preserved. The ideas in this article are described in that filed body, not sketched on a whiteboard and hoped for.

The direction is being noticed. On Crunchbase our founder now ranks number 2, and the company Heat Score reached 94 out of 100, climbing from single digits. We read that as a signal that sovereign, governed autonomy is landing with people who have watched ungoverned automation and understood the bill that comes with it.

Where this goes next

The near term work is to keep widening what the cooperative can safely take on without loosening a single rule that makes it trustworthy. More domain brains, deeper memory that stays owned and layered, and arbitration that scales to busier, more consequential workloads while remaining deterministic and fully recorded. We are not chasing a system that acts more like a person. We are building one that behaves like a dependable institution: capable, patient, and able to prove exactly what it did. That is the version of autonomy an organisation can actually put to work, and it is the one we intend to keep shipping.