Lydia Hallie describes how her team’s coding-agent workflows have changed as models take on longer tasks. She discusses concise instructions and how permission modes can reduce routine interruptions.
Hallie distinguishes model instructions from enforceable boundariesAn agent permission boundary defines and enforces which resources and actions an automated agent is allowed to use.. She explains why deny rules and operating-system sandboxingAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access. address different risks, and notes that tools, hooks and external effects need separate attention.
She discusses verification through tests, hooks and continuous integrationEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria., and describes risk-based review and orchestration of multiple sessions. These practices are presented as her team’s approach and must be adapted to each environment.
Watch on YouTube
