Jim Clark argues that agent safety should come from narrowly scoped capabilitiesLeast privilege gives a person or automated system only the access required for its current authorised work. rather than constant human approval. He separates the agent harnessAn AI agent harness is the software framework that packages a model with tools, instructions, context management, execution controls, and user interaction. from the tools and resources it can reach, using a newsroom workflow and a coding agent's occasional need for commit-signing keys to illustrate task-specific sandboxesAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access..
An MCPModel Context Protocol is a standard way for AI applications to connect to external tools and sources of context. gateway provides a shared control point for exposing tools, resources and prompts across different harnesses. Jim Clark advocates keeping credentials outside agent sandboxes and progressively disclosing only the capabilities required by each taskAn agent permission boundary defines and enforces which resources and actions an automated agent is allowed to use.. He also describes integrating agent authorization with existing organizational identity systems. The talk presents an architectural approach rather than a demonstrated guarantee that agents cannot cause harm.
Watch on YouTube




