I will set up claude code or an ai coding agent with safety guardrails

I set up AI coding agents that do not do damage when nobody is watching.

I run one in production. It goes through job portals every night, scores what it finds, drafts proposals - and sends nothing until I approve it with a one-time code. That approval step is not decoration: it is the difference between a useful agent and one that emails a stranger on your behalf.

What I set up for you:

• the agent itself, with an instruction file written for your repository rather than a generic template
• - explicit boundaries: what it may change, what needs a human, what it must never touch
• - verification that checks the real outcome, not that the command exited zero - a check that only ever passes measures nothing
• - for unattended runs: approval before irreversible actions, plus reporting, so silence never looks like success

I also tell you which parts should not be given to an agent at all. Where a script or a CI rule does the job, that is what you get - it is cheaper and it cannot improvise.

Eighteen years in enterprise IT infrastructure, currently leading a team.

Tell me what you want automated and I will tell you which part should stay with a human.