When the agent is faster at the wrong thing
Agentic speed amplifies the task contract. Define the outcome, protected state, evidence, and stop conditions before optimizing throughput.
Blog
Occasional writing on the ideas and implementation details behind my work.
Agentic speed amplifies the task contract. Define the outcome, protected state, evidence, and stop conditions before optimizing throughput.
A quarterly access review shows why assurance needs a trace from policy clause through assessment activity to evidence and finding.
Legacy data migration works when mappings become versioned, testable artifacts that preserve identity, expose ambiguity, and prove their output.
OWL and SHACL fail when treated as the same governance layer. Let OWL say what follows and SHACL say what passes.
Production ontologies evolve through contracts, not edits. Keep old data intelligible while the model gets sharper.
AI accountability depends on durable decision traces, not role charts or post-hoc model explanations.
AI makes control prose cheap. The hard part is still the evidence chain from risk to test method to durable proof.
Embeddings win at recall over language. Knowledge graphs win when identity, provenance, and policy have to survive change.
Compliance work breaks when evidence lives as static attestation. The useful unit is the edge from obligation to control to proof.
Hardware does not become software when teams add Git-like tooling. The useful transfer is faster, traceable change review.
Approval gates matter when they leave a durable trace: action, policy, evidence, reviewer, decision, and resumed state.
SHACL belongs in production as the executable contract for graph data, with validation reports as reviewable evidence.
RDF and property graphs solve different modelling problems. The useful choice starts with ownership, semantics, and query shape.