Chief of Staff, Part Two: From Jasper to Athena

I wrote about building an AI Chief of Staff a while back, a multi-agent system standing in for a lot of what a real chief of staff does: prep, synthesis, routing work before it reaches me. That system had a name. This is the update, because the honest version of that story didn't end when the first post published.

The original version worked well enough that I started asking it to do more, which is exactly how these things get out of hand. Somewhere in expanding its scope, I lost the thread on what it should be deciding versus what it should only be preparing for me to decide. A chief of staff, human or otherwise, has a very specific job: make the principal's time and attention go further, not quietly accumulate authority because nobody drew the line.

So I built a second profile, a different name, explicitly to test that boundary rather than just extend the first one. The rule I landed on: routing and preparation are the job, not execution of anything consequential. It can draft the options. It can flag what needs a decision and why, with the tradeoffs actually laid out. It does not get to make the call and inform me after, even when it's probably right, because "probably right" isn't the standard for anything that touches money, commitments, or other people.

That sounds like an obvious line until you're actually running one of these day to day, because the whole appeal of a good chief of staff is that they save you from having to think about things. The temptation to let that shade into "and also just decide it" is constant, and it's exactly the temptation worth resisting on purpose.

Where it stands now: a working second profile, a clearer delegation boundary than the first version had, and an open technical issue with one of the routing connections that I'm fixing before I'd trust it on anything time-sensitive. I'll take the L on that in public rather than pretend the sequel is cleaner than the original. It isn't yet. That's the actual update.