Don't let the LLM judge: a Bayesian belief-loop agent architecture proposal

ComprehensiveMonth70 · reddit · 2026-08-18

A Reddit user proposes a brain-inspired agent architecture: instead of trusting a big-prompt LLM to reason, the system maintains competing hypotheses (possible worlds), updates beliefs from evidence via Bayesian posteriors, picks the next check by expected information gain, and chooses cost-aware actions including human escalation.

Using a smart fridge smelling something weird as an example: priors come from door events, temperature history and inventory; after observing 14°C, the system computes likelihoods per hypothesis, updates posteriors, compares entropy reduction, and weighs whether a high-value question like "is the compressor drawing power?" is worth its cost. The LLM extracts signals, proposes hypotheses and calls tools, but belief updates and actions stay with an explicit controller.

Key design points: an "other/unknown world" bucket to avoid assuming a complete hypothesis list; calibrated probabilities with provenance; expected value of information over mere entropy reduction; human escalation for high uncertainty or asymmetric downside; traceable failure attribution. The author asks whether this just reinvents POMDPs, active inference or belief-state planning, and how to handle open-world hypotheses, likelihood calibration, and separating informative questions from decision-improving ones.

Original post →

More from coding & agent

coding & agent channel →