Agents don't fix a broken process. They scale it.
Automating a bad workflow gets you more of the bad workflow, faster and with more confidence. Do the boring diagnosis first.
The first internal agent I built made things worse. It documented meetings beautifully — including meetings that should never have happened, decisions made by the wrong people, and action items nobody owned. The documentation was excellent. The process it described was a mess, and now the mess was legible and permanent.
That is the lesson I keep relearning. An agent is a multiplier, and it multiplies whatever you point it at. Point it at a process with unclear ownership and you get unclear ownership at higher throughput.
So the first week of any agent project is not a prompt. It is a map: who does this today, how long does it take, what makes it fail, and what would we want if we were designing it fresh. Half the time that map ends the project — the answer is a changed meeting cadence, not software.
The other half, you build something narrow that a real team actually uses on a Friday afternoon. That is the bar. Not impressive in a demo — used when nobody is watching.