Don't just say "fix it." Let it run the build, see the error, try again. Agents learn from iteration, not from your first prompt being perfect.
If you explain something to the agent more than twice, put it in a markdown file. AGENTS.md or skills/my-workflow/SKILL.md. Now every agent on your team knows it.
"Fix this specific test" works. "Refactor the whole module" doesn't. Scope your asks like you'd scope a work item. One clear outcome per task.
Agents can use git, az, gh, dotnet, cargo already. You don't need new tools. Just let the agent run what you'd run.
Review the agent's output like you'd review a junior's PR. Don't blindly accept. Don't micromanage either. Find the middle.