I read AI papers for a living, but it has become impossible to track insights.

Every week I triage agent papers for our Top AI Papers of the Week digest.

A few weeks ago, I tried @viktor_com, an AI employee that works inside Slack, in the channel where that triage happens.

I asked it to go through the week's agent papers and pull three worth reading.

It sent the list back that same morning.

On day 19, it posted in the channel without being asked. "Six more landed overnight. One of them contradicts the routing section in module 4."

Two days later, again without me, it drafted the correction and left it unpublished. 2,100 people are on that module.

Everything it did came back as a proposal for me to approve. That is why I kept it running.

Finding new papers is easy. Knowing when one of them breaks something is much harder - in my case, the course lessons I teach.

If you run agents for weeks at a time, look closely at how the approval trail works. Happy to go deeper on this one.