Every model of an agent starts with a boundary. Draw it wrong and the rest of the math does not save you.
Why boundary drawing comes first
An agent, in active inference, is a partition of a joint distribution: internal states that maintain themselves by exchanging influence with external states through a blanket of sensory and active variables (Class E). The Bayesian machinery downstream is indifferent to where you cut. The meaning of the results is not. A partition that lumps the coffee cup in with the person modeling it produces a coherent posterior over something, but not over the person and not over the cup.
Parr, Pezzulo, and Friston (2022) are explicit that this partition is a modeling assumption (Class E). The math tells you the consequences of the cut, not where to make it.
Agency as a statistical claim, not a metaphysical one
The question "is this thing an agent?" reduces, in our reading, to narrower, checkable questions. Does the internal partition minimize variational free energy against a generative model whose sensory and active states are the blanket you drew (Class C)? Does the KL divergence between posterior and prior over hidden states collapse in the expected direction as evidence accumulates (Class C)? Are the policies selected by expected-free-energy minimization reachable through the action channel you specified (Class C)? None of those is a claim about agency in a mind-independent sense.
This puts the burden of proof in the right place. When the partition earns its keep, the belief loop closes on the evidence. When it does not, the partition is the first thing you revise.
Where the boundary lives in a POMDP
In the Precision Lab, the boundary is written into the generative model (Class C). Hidden states are what the agent cannot see. Observations are the sensory blanket variables. Actions are the active blanket variables. Transition and observation models close the loop. That is the boundary.
What we can check, in code (Class C):
- The agent never reads the hidden state. Belief updates are functions of observations and actions only. The Cell Lab falsifier suite surfaces any run where a controller reaches past its blanket.
- The action channel is the only write path. The environment is never mutated except through the specified action variables.
- The posterior is recomputed from the blanket alone. Bayesian inference uses the sensory and action sequences, nothing else.
What this reading does not buy you
This framing avoids several moves we do not want to make. It does not claim that a well-fit generative model is conscious. It does not claim that a Markov blanket certifies a self. It does not claim our partitions are unique (Class U). Two competent modelers can partition the same setup differently and produce internally consistent stories. Only more evidence resolves that.
How the debate lands in our workbench
There is a live disagreement about how much explanatory weight blanket structure can carry, and by extension how much of "agency" it can underwrite (Class E). Our position is recorded in The Blanket Debate, What We Take As Working Assumption, and the minimal mechanics are in Markov Blankets, What They Are, and What They Are Not. We build with the partition, publish it, and let the falsifier suite tell us when it stops paying its way.
A note on Themesis
Themesis maintains a public resource map for entering active inference in 2026. We link it as a factual pointer, not as an endorsement of our claims: Where to Start with Active Inference, A Resource Map for 2026. Our one-line frame, in our voice: the map that lists SWU among five pathways; anchor for family tie-ins.
What the reader should take away
Boundary drawing is the first move, and it is the one most likely to be done silently. Making it explicit turns agency into a set of checkable questions rather than an argument about words. Our partitions, the code, and the disconfirmation criteria are all on the site. If the partition is wrong, we would like to know first.