Not the demo failures. The ones that turn up weeks in, after a client has been using it daily. Reading the threads here on interruptions and handoffs makes me think the demo is the easy part. So for anyone with an agent that's been running on a real client for a few months: what broke that you didn't see coming? Curious whether it tends to be the integrations or the conversation itself.

Quick one for anyone whose agent texts the owner or the caller after a call. In my Twilio logs I found 23 texts that never went out between February and June, while my n8n workflow showed them as sent. All of them failed with error 30044. My account was still a trial, and trial accounts refuse texts that are too long. 10 of them also had a long dash, and one character like that makes each part of the text hold 67 characters instead of 153. For an agent whose whole job is to text the owner about an urgent call, a text that fails without anyone noticing is about the worst bug it can have. So my templates use plain characters only now, and I check the delivery status in Twilio rather than trusting the workflow. Has anyone else found failures like this long after they happened?

I spent some time on a plumbing forum to see how owners actually talk about missed calls. One thread stuck with me. A guy running a 3-person shop posted 47 missed calls in a month, only 5 voicemails. He'd already tried his wife on the phone and a part-time receptionist at $1,500 a month. The plumbers who answered weren't impressed. One misses 5 to 10 calls a day and says most are price shoppers. Another said to tag every missed call for 2 or 3 weeks before buying anything, and when his shop did that, most of the calls were junk. From now on, when I talk to an owner, my first question will be how many of his missed calls were real jobs, before I mention the agent at all. If he doesn't know, that's the first thing we find out together. How do you handle owners who say "most of those calls are garbage anyway"?

AI now writes about 90% of Claude Code's own code. So who reviews it all? Not one human reading line by line. That stopped working as a gate a while ago. Anthropic just published how its team replaced it. The results from June to September: → Merged PRs more than doubled → The share of PRs needing a revert fell by more than 40% → PRs passing after one review round went from 65% to 78% What changed was the review system, not the reviewers. It works in six layers: 1- Verify before the PR exists. The agent runs tests, build and lint and shows the output, instead of just saying it worked. 2- AI review on every PR. Several agents look for bugs in parallel, and the findings get voted on before any are posted. 3- A security review that can block the merge. 4- Required CI. 5- Human approval scaled to the risk: 0, 1 or 2 people depending on what the change touches. 6- Feature flags and staged releases, with one-step rollback. The rule I'm copying first: cap review rounds at 3. Agents will happily fix each other's nitpicks forever. One PR ran more than 50 overnight rounds before its author stopped it. What humans still own: judgment, direction, the risky code paths, and the outcome of every change that ships. The full breakdown is in the carousel. Anthropic's full write up is in the comment.