Unsupervised Agent Discovery
Recover agent-like boundaries from timestamped state-variable traces without pre-labeling which variables belong together — using conditional-independence cuts, lagged memory analysis, and intervention handles when passive statistics are ambiguous.
What decision changes?
Do not take variable names, vendor rosters, or subsystem maps as the agent census. Run an unsupervised partition on the traces first, then audit the units the method infers.
Boundary discovery asks where the operative controller is. Unsupervised Agent Discovery (UAD) is one operational answer: given timestamped traces of many state variables, infer which subsets form coherent agents — private memory, sensory interface, action channel, and approximate conditional independence from the rest — without taking labels, vendor rosters, or subsystem maps as ground truth.
The pipeline follows Markov-blanket logic. Partition variables into internal, sensory, active, and external roles; test whether the sensory–active interface screens inside from outside; extract memory substrates by lagged mutual information; and, when passive statistics are ambiguous, use intervention handles to see whether disabling a candidate channel changes behavior more than ordinary run-to-run variation would predict. That last step is intervention-supported unit discovery , not a separate problem — it is what UAD looks like when correlation alone merges an entire pipeline into one false unit.
UAD is the measurement bet behind MB1 (boundary estimator soundness): if the procedure issues an ε-boundary certificate from traces, later alignment steps inherit whatever error the partition missed. Downstream uses include inferential-coupling detection over discovered agents (Ch. 35 ) and census questions in selection environments (Ch. 34).
Claim strength is deliberately mixed. Controlled simulations recover known partitions; adversarial and noisy substrates expose when cheap heuristics suffice, when passive clustering over-merges, and when shared observation channels smooth per-agent signal beyond what pairwise tests can resolve. The negatives are kept alongside the positives.
Experiment lines that stress UAD
| Line | Card | Findings |
|---|---|---|
| Agency-detect (sibling) | Experiment card | Findings |
| Embedded audit simulation | Experiment card | Negative ledger |
| Lab-layer simulation | Experiment card | Findings |
| Graded-capability lab | Experiment card | Findings |
Start with agency-detect for the original prototype, or embedded simulation for the richest end-to-end negative ledger (ES-1: UAD equals loudest-actor heuristic in the default ecology).
What would count as evidence?
Partial support from controlled multi-agent simulations and lab substrates with known ground truth; extensive recorded negatives when heuristics suffice, observation channels smooth per-agent signal, or passive clustering merges whole pipelines.