Developer Anthony Ettinger says one short request to an AI coding agent, a link to his GitHub issue list with a question asking for fixes, ran his whole morning. His AI agent coding workflow left him approving steps while the software did the engineering.
By midday, he says, the agent had closed six issues, merged four pull requests and pushed three npm releases. It also upgraded and verified three production servers and shipped a Next.js security patch. Ettinger typed no code.
What got done before lunch
A payment bug that had quietly broken PayPal checkout for 47 accounts got fixed too, with a data migration and a live test order as proof. On his desktop, two background services had been crash-looping about 450,000 times between them, and the agent cleaned that up as well.
His own part was brief: solve one CAPTCHA, follow a magic link from an email, paste in an API key, and answer prompts with words like “do it” and “check now”. When PayPal failed, he told the agent to fix it because checkout was unusable.
Where the human was still needed
A threat intelligence provider had ended its free downloads and now required a free API key. The agent wrote the integration and confirmed the endpoints with a dummy key, then halted. Registration sat behind a Cloudflare Turnstile check, and it refused to get around it.
The agent pulled the login link from its own inbox and signed in, but a second CAPTCHA blocked key creation. It passed Ettinger a link with instructions, he made the key and handed it back, and the agent rolled it out to each server. Indicator counts jumped from 5 to thousands, and the agent noted the key’s location for future sessions.
More than faster typing
Ettinger describes work that goes beyond autocomplete:
- Five of the six open feature requests turned out to be shipped already. The agent read the merged pull requests, worked out what each still lacked and built only that.
- It handed pieces to parallel sub-agents working in separate git worktrees, reviewed what came back and fixed the one branch that clashed. Fresh eBPF code got a nine-second run as root on a live kernel before merging.
- One production box had rolled itself back to an older version since an install directory had belonged to root after an incident response two days before. The agent checked the incident notes and judged it harmless.
- A download kept breaking partway through. Testing from another machine showed the upstream server was responsible, so the agent added retries and a one-day fallback.
Judgment calls stayed with the human
PayPal checkout kept dying because each deploy locked the merchant’s encrypted vault again. After a wallet was drained the previous month, Ettinger had banned any server-side key that could decrypt user secrets. The agent respected that, explained the tradeoff and waited. He then said human sign-off should be needed only when truly critical. The agent applied that solely to credentials that take in money, left spending credentials untouched, logged the decision as a permanent rule and shipped.
Other calls came from experience too. His antivirus once quarantined what appeared to be a reverse shell. Ettinger wanted to know its origin before wiping anything, and it proved to be a test string from a security tool, sitting in a 20,000-skill archive he had cloned in bulk months back. For a failing gateway whose package had disappeared, the agent skipped the obvious reinstall, since that framework belongs to the same ecosystem as the archive of malicious skills. It spotted a remote-control service open on all network interfaces, guarded only by a URL token, and Ettinger had it restricted to his private network. Before removing a 5.6 GB repository, it looked for unpushed commits.
Mixed feelings about the shift
Ettinger says he missed the pleasure of debugging. One systemd unit had restarted 280,589 times because an install script left a blank path in ExecStart, so it kept launching node //src/index.js. He got a summary rather than the hunt.
He still rates the outcome as good, with patched servers, working checkout and populated threat feeds. His value, he says, has shifted from knowing how to do a task to knowing which tasks matter and which risks to accept.
How he works now
Ettinger now keeps several agent sessions going in the background and checks on them. He records decisions for the agent to reuse and spends more time on policy, such as which jobs must stay manual, than on code. For people starting out, he recommends operating live systems and reading postmortems, since that experience teaches which buttons to press.
Source: dev.profullstack.com