Identity, tools, memory, software supply chains, and the operational failures that appear when models can act through real infrastructure.
Keep a steady time counter (TAI) apart from a position in the solar system. The split makes replayed agent memory match byte for byte—within limits set by astronomy and metrology.
Agentic risk awareness is an operational discipline: model the failure paths, constrain authority, force uncertainty to surface, and continuously test whether the system still behaves under stress.
Every third-party tool an agent invokes is someone else's code running near your credentials.
Chain-of-thought traces can be inspected as dependency graphs instead of prose, making circular support, abandoned threads, unsupported conclusions, and contradictions easier to find.
Agent systems need a memory control plane because retrieval safety and historical truth require enforceable policy and clear lineage, not mutable notes.
A genetic prompt run improved average fitness from 0.887 to 0.926 in four generations using LLM-based mutation and crossover.
Send notes on an article, questions about the research archive, or a failure mode worth examining. Stack Research is a small independent effort; useful disagreement and field evidence are welcome.
