Reflective Labs / Notes
Why context is more important than ever
When agents talk to each other directly, the conversation becomes the architecture. When they only read and write a shared context, the context becomes provable.
The instinct in almost every early multi-agent system is the same: let the agents talk to each other. One proposes, another responds, a third arbitrates. It feels natural — it's how people coordinate — and it is, we think, the single most consequential architectural mistake available in this field, because it works just well enough to fail quietly and expensively later.
When agents address each other directly, three things happen. Hidden coupling appears: an agent's behavior now depends on who spoke to it, in what order, and with what phrasing — none of which is visible from outside the conversation. Explainability collapses: the question "what facts led to this decision" becomes "which agent said what to whom, and when," which is a much harder question to answer months later. And evaluation becomes brittle, because the unit under test is a conversation rather than a function of state.
Our position is that the context — the shared, append-only set of facts a Formation is reasoning over — should be the *only* channel. Suggestors read it and propose to it; they never address each other. This sounds like a modest engineering preference and it is actually load-bearing for the mathematics: if one Suggestor could message another directly, the order in which they ran would be observable in the result, and the commutativity invariant our determinism proof depends on would be false by construction — not hard to enforce, but meaningless to even state. Because context is the only channel, each Suggestor's contribution is a pure function of context alone, and commutativity becomes a property you can require, check, and rely on.
This is also why context is becoming more important rather than less as models get better. A more capable model does not reduce the need for a shared, attributable record of what is known and by whom — it increases the stakes of getting that record wrong, because a more capable model will act on it faster and with less visible hesitation. Context is not a convenience feature bolted onto an agent loop. It is the API, in the strict sense: the only interface through which participants in a Formation are allowed to affect each other, and the reason a fixed point can be proven to exist at all.