Agent Heartbeats
Continuing from , I figured I should reflect a bit on using what I'm referring to as "agentic heartbeats". The gist is setting up a scheduled task in your harness of choice, and having the agent start with a common prompt to do some work, mostly in my case that work is research.
Specifically I started with an hourly schedule (I’ve since shortened it to 15 minutes) and had the agent look at my current work (threads in my harness), any inbound signals like GitHub notifications or slack notifications, newly created notes in my knowledge base application, and anything else we tabled to return to, it digests all this information and then sends a high level summary to my orchestrator thread.
So far this has been pretty useful (I’ve had my agent successfully triage bugs I’ve saved in my notes and open up PRs to fix them), but it also seems to communicate the last mile of things that need my human attention. For example it’ll often say something like:
“The package integrity failure has now affected 17 repositories“
I have no clue what my agent is talking about there, what is the "package integrity failure"? What are the 17 repositories? Why should I care?!
At the moment the next thing I want to think about is some form of shared inbox, a way for an agent (either through this heartbeat setup, or ad-hoc) can send me notifications (ideally with well explained summaries!)