Something I noticed running an automation agency Maybe the GTM guys can validate this. Best lead magnet for an automation agency: 1 Pick a business. 2 Map how it runs. 3 List what you'd automate and why. Send it for free. "We do AI automation" gets ignored. A map of their own workflow gets read. Of course you cannot really find everything about them online.

most automation requests I get aren't ready to be automated. not because AI can't do it. because nobody can explain how the job gets done today. if you can't write the steps down, a bot can't follow them either. fix the process first. then automate it.

my AI assistant told me the job was done. it wasn't. it wasn't lying. it told me what it meant to do. if you automate part of your business, don't trust the "done" message. check the result. every time, until you have a reason to stop.

10/10 plot twist

Before I switch on automatic deploys to anyone's production, four questions: → which branch deploys → auto on merge, or manual trigger → where do failed deploys get reported → is main protected Skip the fourth and any direct push ships to production.

Agents are not replacing people. People using them are replacing people who don't. That's a much less comfortable sentence and it's the one that's actually happening.

Needed a real Postgres to test SQL. Docker was down. initdb a throwaway cluster, pg_ctl on a spare port, ten seconds. Gotcha: a socket path inside a temp directory blows past Postgres's 103 byte limit, so force TCP with -k ''. Faster than fixing Docker.

macOS sed has no word boundary. GNU sed treats \b as zero width. BSD sed treats it as the letter b, matches nothing, and exits 0. So your substitution silently did nothing and told you it worked. On a Mac, assume any GNU regex extension no ops instead of erroring.

Wrote a safety check comparing the SHA I tested against the SHA about to deploy. It printed "same" for every branch. I nearly trusted it. zsh does not word split an unquoted variable, so both sides were empty strings. A check that compares nothing to nothing always passes.

Our staging environment was running a build from three months ago. Same service name, same dashboards. 70,000 proxy errors in 48 hours from credentials nobody had used since June. Anyone who looked at staging concluded the service was broken. It wasn't. Staging was.

I reported 481 rate limit errors in our logs. Then I retracted it. The "429" was matching inside lines like Duration: 2429 ms. There were no rate limit errors at all. Search logs for phrases. Never for a bare number.

Wrote a job that gives up after five tries and sends nothing. Not a bug. A daily digest arriving six hours late is worse than one that never arrives, because you read it as current. Half of alerting design is deciding what not to send.

Wanted a daily report that sends exactly once, through restarts and two replicas. Didn't build a scheduler check. Built a row. Insert the date, let the unique constraint reject the losers. The insert is the lock. Exactly once is a claim someone wins, not a clock you trust.

Found a bug that flips a finished job to failed. The handler skips a task that's already running. Its finally block still writes finished, then the default status, which is failed. No state check. So one queue redelivery marks a successful job failed. Retries are only safe if the writes are.

Qwen3.8-35B-A3B-Distill - our first MoE. Qwen3.8 reasoning distilled into Qwen3.6-35B-A3B: 35B total, ~3B active. ARC-Challenge 0.548 → 0.591. MMLU holds at 0.834. Training data partly from our free community endpoint, thank you! Weights + GGUFs: huggingface.co/empero-ai/Qwe…

Two pull requests. Git says they don't conflict. One adds a guard that rejects any non numeric id. The other changes ids to text so it can accept UUIDs. Merge them in either order and the second one silently deletes the first one's point. Git diffs lines. It does not diff intent.

A test hotel in our scraper suite is seasonal. Ten months a year it returns sold out. Sold out counted as a pass. So the test was green and measured nothing, for months. A passing test is not a working test. Check what it asserts when the data is empty.

Spent a day proving a scraper was fine. 500 calls in production, 500 locally, five countries. 100% both. Then I read the code properly. The endpoint I was hammering never touches the database. The queue path does, and that's where the failures were. 100% pass on the wrong path.

Automatic deploys are not the hard part. Pull and restart is twenty lines. The part that makes it safe: record the current SHA first, health check after, reset back if it doesn't come up. Most people automate the happy path and find out on a Friday.

Shipped a deploy pipeline for a client this week. Amount of AI in it: none. A systemd timer every 60 seconds. git fetch. If the SHA changed, rebuild, restart, health check. That's it. Most of what people want an agent for is a cron job and an if statement.