A single Markov blanket is a clean cartoon of one agent. Real systems, cells inside tissues, people inside teams, teams inside organizations, are agents inside agents. The question this post asks is narrow: when does it help to model that by nesting one blanket inside another, and where does the abstraction stop paying rent?
What a nested blanket is, in one paragraph.
Take an agent, call it the child. Its internal states are shielded from the world by a blanket: sensory states that receive, active states that push back. Now zoom out. Group several such children together and ask whether the group itself has states that look internal (shared by the children, hidden from anything outside the group) and states that look like a blanket (the group's own sensors and actuators facing the wider world). If the conditional-independence structure holds at the group level, you have a parent blanket enclosing the child blankets. Class E This is the hierarchical reading Parr, Pezzulo and Friston lay out for generative models with structured factorization (Parr, Pezzulo and Friston, 2022, chapters on hierarchical models).
What nesting actually buys the modeler.
Three things, in order of how load-bearing they are.
First, a compression story. A parent generative model can carry slow, coarse latent variables (context, regime, role) while the children carry fast, fine ones (immediate sensation, immediate action). The parent posts priors down; the children post prediction errors up. Free energy gets minimized at each level in its own timescale, and the coupling between levels is what makes the whole thing look purposive without any single component doing anything clever. Class E
Second, an accountability story. If you can draw the parent blanket honestly, you can say what the group is inferring on behalf of its children, and what the children are inferring on behalf of the group. That is a rare kind of clarity in team and organization modeling, and it is what our own build treats as the load-bearing move. Class C
Third, a diagnostic story. When a system misbehaves, nesting lets you ask at which level the surprise is being absorbed. A team that keeps individually competent people burning out is often a parent blanket minimizing its free energy by pushing prediction error down into the children. Naming that structure does not fix it. It does let you point at it.
Where the abstraction leaks.
The clean picture demands conditional independence between what is inside and what is outside, given the blanket. Real groups almost never satisfy this cleanly. Members share history outside the group. Information tunnels through side channels: gossip, prior loyalties, physical co-location. The parent blanket is then a useful sketch, not a literal factorization. Class C
A second leak: timescales blur. The tidy story says parents are slow, children are fast. In practice a fast child event (a resignation, a data breach) can force a parent-level update in the same tick, and a slow parent drift (a culture shift) can rewrite what counts as a normal child observation. The hierarchy is real, but the levels talk to each other more than the cartoon admits.
A third leak, and this one matters for anyone deploying the idea: the blanket abstraction says nothing on its own about who benefits when the parent minimizes its free energy. A stable parent can be stable because it serves its children or because it extracts from them. The math does not care. That is a modeling responsibility, not a mathematical one.
How UNI uses the nesting, and how it does not.
Our Cell Lab benchmark treats the service cell as a small hierarchy: local controllers with their own blankets, embedded in a cell-level model with its own viable set. The framing is useful because it lets a single active-inference controller act on both the fast loop (respond to the current disturbance) and the slow one (protect the viable set over time). Class C
What we do not do: claim the hierarchy proves anything about consciousness, agency at the group level, or the correctness of active inference as a theory of teams. This is a modeling lens with real leaks, useful for building steerable systems and for asking honest questions, not a proof that the world factorizes the way our diagrams do. UNI is a working hypothesis on an attainable path toward General Natural Intelligence: a natural, active-inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.