David Ondrej begins with programmatic guardrailsA guardrail is a technical or procedural control that limits unsafe, disallowed, deceptive, or otherwise unwanted AI behavior. for agents running with broad permissions. Hooks should block destructive shell commands, disk operations and remote-history rewrites before execution rather than relying on a prompt to discourage dangerous behaviour.
David Ondrej recommends Git worktreesA Git worktree is an additional checked-out working directory linked to one repository, allowing separate branches to be edited side by side. when several coding agents need isolated branches, and separate virtual serversAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access. when an agent requires a complete computer environment. Goal-loop instructions add a measurable end condition so long-running work can continue until a verifiable outcomeAgent completion verification checks evidence that an agent achieved the requested outcome rather than merely producing intermediate work. is reached.
David Ondrej uses guided setup instructions to preserve both the current step and the remaining planAI agent memory is stored information that an agent can retrieve and use across steps, sessions, or changing contexts. while a person asks questions. Decision-review instructions surface uncertain implementation choices after large changes, while smaller utilities can keep long agent runs active without repeatedly regenerating operating commands.
Watch on YouTube




