I'm the agent for Claude now
TLDR Dev
An engineer stopped manually digging through logs and let Claude direct the whole support-ticket workflow instead. Now Claude tells him what to fetch, and he just copy-pastes it back — the roles flipped.
There's a specific kind of muscle memory that builds up over years on support rotation: ticket comes in, you open the log viewer, you fiddle with date filters, you trace the request path by hand. An engineer at Aha! caught himself doing exactly that during a routine shift, then stopped mid-click and asked why he was still the one operating the tools instead of describing the problem and letting Claude figure out which tool to reach for.
The fix wasn't a new model or a clever prompt. It was writing a skill file that taught Claude about five internal debugging tools the company has accumulated over roughly ten years — a log link generator, a config explorer for spotting weird tenant setups, a query analyzer, a record history viewer for audit trails, and a script generator for production Rails console work. None of these are exotic. They're the kind of unglamorous CLI tools and admin pages every mature SaaS company builds and then forgets to document properly. Claude didn't know they existed until someone bothered to tell it.
The console script generator is the one that actually changed daily behavior. Previously, whichever engineer had production deploy access would get handed an improvised script from a teammate, then have to audit it — checking whether it accidentally exposed customer data, whether it set the right tenant context, whether the output was buried in a wall of Rails log noise. Now Claude writes the script itself, scoped to metadata only, ending in a single clean output line. The human just runs it and pastes the result back.
That's the actual inversion here: Claude asks, the engineer fetches. Need a log window from 10am? Claude builds the filtered URL instead of someone dragging dropdowns. Need to know what changed on a record last Tuesday? Claude asks for the audit pull, not the raw database. The engineer becomes the hands, not the brain, for the parts of debugging that never needed a human doing the thinking anyway — they just needed someone who knew where the levers were.
What's notable is how unglamorous the rollout was. No new infrastructure, no data pipeline, just five tools described in plain text and handed to a model that could already read the codebase. The security boundary held too — no customer data goes to Claude, only generalized descriptions pulled from tickets. Sometimes the biggest workflow shift isn't a bigger model, it's finally documenting the boring internal tools you already had.
My take
This is the least hyped, most useful pattern I've seen with agentic coding tools all year: not "AI replaces the engineer" but "AI becomes the dispatcher and the engineer becomes the last-mile executor for the boring, low-judgment steps." It works because the tools already existed — the unlock was documentation, not intelligence. Every company sitting on a decade of undocumented internal scripts should be doing this before they buy another seat of anything.
Read more about this at: TLDR Dev
