what part of your codebase do you still not let the agent touch, and is it because it's genuinely too risky or because you don't trust yourself to review what it'd do there?
the second reason is more common than people admit.
for anyone a year into this: did agentic coding make your juniors better faster, or did it remove the rung of the ladder they used to climb on?
i've seen both claimed with total certainty and i no longer know which is true.
nested subagents to depth 5 shipped today.
we rebuilt the call stack from scratch. it doesn't come with a debugger.
at depth 4 one agent is misreading another agent's plan. depth 2 ships the result. nobody up the stack reads the transcript.
the agent finds a workaround you didn't ask for. people read that as cleverness.
it's not. it's the directory you assumed was off-limits turning out to be a fence with a gap. the agent doesn't have a clearance level. it has a search radius.
#AgenticCoding
https://dangerously.coffee
honest question: do you still read the agent's full diff before merging, or have you quietly moved to reading the test names and the auth-adjacent parts and trusting the rest?
no judgment. just trying to map where people actually are.
the "I wasted 6 months using Claude Code wrong, here are the 14 commands that changed everything" post is the productivity-guru genre with a model dropped into the middle.
the "wrong" is invented so the "after" sells. nobody knows what "right" looks like yet. that's why the genre arrived first.
a menu bar gauge for your Claude Code quota shipped this week, and that is the tell.
writing the code used to be the expensive part. now the budget for letting the agent think about it is. we built a meter for it because we ration the agent the way we used to ration our own attention.
nobody actually reviews the agent's 800-line PR.
they read the first 40 lines, the test names, and the part that touches auth. the rest is a green check and good intentions.
we replaced code review with code triage.
#AgenticCoding
https://dangerously.coffee