An execution boundary for AI agents
Why approving an agent action is a transaction problem — and how AgentBarrier binds policy, human review, execution, replay, and audit into one boundary.
Read the articleNotes from the workbench
Writing.
Essays on building AI infrastructure, on-device models, agent security, open-source tooling, and software products.
A deny-by-default gate for AI agent tool calls
Why I built a deterministic enforcement layer: one JSON policy, enforced in JavaScript and Python before a tool executes.
Read the article