Writings

A framework for building harnesses around LLM-powered agents through deliberate constraints, external validation, and composable tooling.

Why "memory" is the wrong abstraction for long-horizon agents — staleness, lost provenance, timeless edges, conflated stores — mapped to the prior research that has already isolated each failure mode.

Saliency vs. extensionality — why output-length bias, attention dilution, and decode economics push agents toward salient answers when extensional ones are required, and a sketch of a typed operator language built on Structured Agency as the principled fix.