Focus
Agents
Multi-step systems where the hard part is handoffs, state, and failing closed — not a clever prompt.
Agents fail at orchestration, not reasoning. A chain of five 90%-accurate steps is a 59% system unless every boundary is boring and checkable, and most of that loss is unchecked error propagation rather than model capability.
So the work here is mostly about the seams: what a step is allowed to hand the next one, what happens when the evidence is weak, and how a system says it does not know instead of guessing.
On this shelf
HF Agentic SearchWork · GitHub + live SpaceA research loop that plans searches, reads the evidence, notices when it is thin, and searches again rather than returning the best of a bad set.OpenCompliance AgentWork · 9 test modules + CIExplainable YAML rule packs over RAG retrieval — screening, not unsupervised legal advice. Backed by a golden-sample test in CI.OpenAgents fail at orchestration, not reasoningNotesWhy multi-agent accuracy compounds down, and what to instrument first.OpenDeterministic gate for multi-agent validationDesigned, not builtSchema validation and state reinjection between every agent handoff. A design; nothing implemented.Not builtGraph-augmented agentic RAGPrototype, no repoWhen vector search is the wrong topology and the graph is the unit of reasoning. Exercised on a synthetic corpus only.Not built