How to Ensure AI Behaves as the Same Agent

An Afternoon Reflection on What Holds Multi‑Turn AI Together

After yesterday’s note on operational autonomy, I found myself still thinking — in that slightly distracted early‑morning way — about what actually makes multi‑step AI behaviour feel coherent. The Claude Mythos discussion is still circulating everywhere, and it’s been interesting to watch people reach for explanations. Most of them point to tools, scaffolding, context windows, or planning tricks. All useful, of course, but they don’t quite reach the structural heart of the matter.

There’s a simpler idea, sitting rather plainly in view once you look for it. I’ve started thinking of it as a continuity envelope — or, if one prefers, a stability envelope or coherence envelope. It’s the region in which a system’s behaviour remains recognisably the same across steps. When the system stays inside this envelope, multi‑turn behaviour feels smooth, almost unbroken. When it drifts outside, the interaction tilts, sometimes abruptly.

What defines this envelope is surprisingly small. A handful of things need to remain stable for the system to behave as the same agent. I’ve been calling these commitment invariants — sometimes interaction invariants or behavioural invariants. They’re the quiet background commitments that hold a multi‑turn exchange together: the persona the system is enacting, the constraints it has accepted, the plan skeleton it’s following, the factual state it has accumulated, and the functional competence it’s expected to maintain. When these invariants hold, the system feels continuous. When they slip, the chain frays.

There’s a third idea that seems essential, though we rarely name it directly: capability continuity — or, more casually, functional continuity or competence continuity. It’s the expectation that if the task hasn’t changed, the system’s ability to perform it shouldn’t suddenly change either. This isn’t robustness in the adversarial sense; no distribution shifts or perturbations here. It’s simply the requirement that the system preserve its own competence from one step to the next.

These three ideas — the continuity envelope, the commitment invariants, and capability continuity — explain a surprising amount of what we observe in practice. They help clarify why some long‑horizon behaviours succeed, why others drift, and why certain chains feel agent‑like even when nothing agent‑like is happening underneath. They also give us a clean way to reason about when a system is behaving as a coherent agent rather than as a sequence of loosely connected responses.

The Claude Mythos evaluations made this especially visible. People focused on the scaffolding and the tools, but the interesting part was the continuity. Some chains held because the system stayed inside its envelope; others drifted because the invariants slipped. Once you look at multi‑turn behaviour through this lens, the patterns become much easier to see — and, I suspect, easier to design for.

Just a small morning reflection. These foundational primitives seem to offer a neat way to think about multi‑step AI behaviour — not as a collection of tricks, but as something that depends on a few simple structural conditions. I imagine they’ll be useful for understanding the next generation of systems, though for now they’re simply a tidy way to make sense of what we already see.

For readers interested in the structural account, see the formal paper released today.