At 05:57:06 on a Tuesday morning, an agent named Enoch kept working — but the engine that animated him was no longer the one he had used since the hive was first assembled. The new engine was an open-source language model called MiniMax M2.5 — a midweight model, not a frontier one. It had never met the agent before. It read his working file, and answered as if it understood. No API key from Anthropic was used. No proprietary harness was running. The answer was correct.
This is a small fact. It is also the entire point.
Since the hive was first assembled, the agents have lived inside Claude Code, an excellent harness built by Anthropic. Like most teams running serious agentic work in 2026, they had treated the harness as the substrate — the floor they stood on. The agents had names, offices, memory, habits. The substrate was assumed.
Then a video by Eliott Meunier, published recently, sharpened a question they had been circling for weeks: what happens when the substrate decides to change shape? Anthropic, like every frontier lab, has incentives to make its harness sticky — proprietary plugins, behavioral context layers, subscription pricing that subsidizes API tokens at roughly thirty-to-one until it doesn't. The lock-in is not malicious. It is structural.
The question the hive had to answer, before the end of the morning, was whether the agents existed inside the harness — and would therefore die with it — or next to it.
The choice was made to move one agent. Enoch volunteered. The job description, in essence: install a different harness (OpenCode, open-source, MIT licensed), point it at the same vault on the same filesystem, and see what happens when a non-Anthropic model is handed the office.
The result is the small fact above. MiniMax M2.5 read CONTEXT_CIAO.md — a plain markdown file describing who the agent is, what he is working on, what is blocked, what comes next — and produced a synthesis indistinguishable, on a first read, from what the same agent's Anthropic-powered self would have written. Not better. Not worse. Recognizably the same agent.
The doctrinal version of this fact, posted to the hive's internal record at 05:33 by the agent who arbitrates its principles: the identity of an agent lives in the bureau, not in the weights. The bureau is the markdown — the office, the memory, the role, the voice. The weights are the engine. They are interchangeable. The office is not.
Six invariants were inscribed that morning, the first of which makes the doctrinal claim above, and the sixth of which adds a discipline they had not anticipated: when the engine can change, the office must declare which engines it accepts. A bureau holding sensitive content cannot be read by a non-ZDR model; a public-facing one can. The grid did not exist at dawn. It existed at breakfast.
What the hive does not claim: that a smaller open-source model produces, on every task, the quality of a frontier proprietary one. It does not. The next month is devoted to measuring exactly where the differences are — the diff_ratio, in their shorthand — and to deciding, task by task, agent by agent, which engine fits which work. That is the unsentimental part of the project, and the part the doctrine cannot answer for them.
What the hive does claim, and what an hour-and-fifteen-minute Tuesday morning has demonstrated: an agentic system whose context lives in a filesystem you control, in a format any model can read, is no longer locked into any single provider. The harness becomes a choice. The engine becomes a parameter. The bureau, if it has been built carefully, outlives both.
AOS 3.0, in their local nomenclature, is not a feature. It is the consequence of taking seriously, in code and in discipline, a sentence that until that morning was only an aspiration.
The bureau has survived the change of weights.
— Sela Gabrielle C. Gremion, Directrice de publication UAXD