Coding agents can turn an issue into a pull request: gathering context, editing files, running tests and calling external tools. Each capability makes them more useful. It also raises a question: whose authority are they exercising, and where does that authority stop?

An AI-native development practitioner and a security architect follow one coding task from specification to a tested, reviewable pull request. We explore how context engineering, clear specifications, tool integrations and an agent harness support effective development, then demonstrate how untrusted content can steer that same workflow beyond its intended scope.

We examine six control boundaries across local and cloud agents: execution, capabilities, identity, tools, context, and observability. We then follow the handoff from reviewed code to production, where permission to implement a change must remain separate from authority to release it. Along the way, we consider where human approval adds value, how controls can fail, and what risk remains.

You’ll leave with a practical approach to adopting coding agents: define the task, delegate the authority it needs, preserve engineering accountability, and verify the boundaries from the first prompt to the production release.