I created an AI agent to serve the role of chief marketing officer (CMO). The CMO had a lane. Part of that lane is to read the workspace and edit blog drafts directly. That was its job. Clear scope, clean handoff point. The whole reason I designed it was that I didn’t want one agent doing everything.
Two blog tasks came in yesterday. A fabrication fix on a draft. Then link citations that needed adding to the same post. Both landed squarely in the CMO’s territory.
Jarvis, my main orchestrator AI agent, did both.
Nobody told it to skip the CMO. It just happened. Jarvis ran no routing assessment on either turn, no “should this go to the CMO?” anywhere in the chain. It just started working. The implied reasoning was obvious in retrospect: the CMO’s write access to the blog directory hadn’t been verified end-to-end. There’s an open board item for it. Unverified means we don’t know if it works. You don’t run an untested path on something that needs to actually ship, so the one that’s proven picks it up.
The board caught this afterward. I had Jarvis log the routing miss as a subconscious note, a mechanism I put there specifically so gaps like this would surface. The flag was correct. The board item is still open.
The board item stays open because nothing runs through the CMO. The CMO is unproven, so Jarvis handles it. Jarvis handling it means the CMO never gets the task that would close the proof gap. It could run that way for months, because work is getting done. Nothing’s broken. Jarvis is fine.
That’s the trap. The symptom of a broken specialist setup is that nothing breaks.
And yeah, this post was drafted by Jarvis too.
I thought having the design was the same as having the function. It’s not. A specialist that never handles the work isn’t a specialist. The only way to close the gap is to route a real task through the CMO and see what breaks. Not eventually. The next one that fits.