A Markov blanket is not a wall the universe hands you. It is a set of variables you decide to treat as the interface between what you call "inside" and what you call "outside." Change the decision, and you change the agent.
In the Parr, Pezzulo and Friston formulation, a Markov blanket is defined by a conditional-independence structure: given the blanket states (sensory plus active), the internal states are independent of the external states. (Class E, Parr, Pezzulo and Friston, 2022.) That is a mathematical property of a specific partition of a specific graph. It is not a physical membrane, and nothing in the definition tells you which variables to put on which side.
So a modeler picks. The picking is the modeling choice, and it is where most of the interesting disagreements about "what is an agent" actually live.
Same system, three legal partitions.
Take a person answering an email. Consider three ways to draw the blanket.
- Partition A, the skin. Internal states are neurons and viscera. Sensory states are afferent signals at the sensory surface. Active states are motor outputs. External states are everything past the skin, including the keyboard and the email.
- Partition B, the workstation. Internal states include the person plus the machine plus the drafts folder. The blanket is the outgoing send button and the incoming inbox. External is the network beyond that.
- Partition C, the team. Internal states are a whole small group and their shared documents. The blanket is the group's public commitments and its public inputs. External is the market they operate in.
All three partitions can satisfy the conditional-independence definition, if the variables and conditional structure are specified consistently. (Class C.) Each one gives you a different agent, with different beliefs, different actions, and a different free-energy landscape to minimize. None of them is "the true boundary." They are three usable models with three different jobs.
The choice is disciplined by the question, not by taste.
Which partition is right? The one whose internal-state posterior best serves the question you actually want to answer.
- If the question is about attention or perception under noisy input, Partition A (the skin) tends to be the useful cut, because the sensory states are where precision-weighted prediction error lives.
- If the question is about a knowledge worker's throughput, Partition B (the workstation) is often the honest cut, because the drafts folder and the machine are load-bearing parts of the inference. Treating them as external throws away most of the signal.
- If the question is about a team's decision quality, Partition C is the cut that matches the phenomenon. Trying to explain team decisions from the neurons of one member is a category error.
None of these is a metaphysical claim about where the self ends. They are claims about where to cut the graph so the model does useful work. That is what we mean when we say the boundary is a modeling choice.
Two practical consequences.
First, "does this system have a Markov blanket" is almost never the interesting question. Most systems admit many partitions that satisfy the conditional- independence property. The interesting question is which partition you are committing to, and why. State the partition explicitly and the argument becomes tractable. Leave it implicit and it becomes a debate about intuitions.
Second, different partitions give you different agents with different responsibilities. In our own applied work on service cells and treatment loops, we sometimes model at the box level and sometimes at the loop level, and the inferred "agent" is different in each case. (Class C.) A control policy that looks optimal at the box level can look pathological at the loop level, and the reverse. The math is consistent. The moral of the story is that you have to pick a cut and be honest about it.
What this is not.
This note is not a claim that boundaries are arbitrary. Conditional independence is a hard mathematical constraint: not every partition of a graph will satisfy it, and the ones that do are constrained by the causal structure of the system itself. The freedom is in which valid partition you select, not in inventing partitions that violate the graph.
It is also not a claim that we have decided which cut is right for consciousness, agency, or personhood. 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.