I've been trying to get the most out of my personal AI licenses, so I've been doing this thing whenever I'm about to step away from the computer for awhile. I open a fresh terminal, give the agent a long running task, and leave.
The bigger the task, the better. Bonus points for fanning out subagents and using feedback loops. When I return, I have something new to check out. Sometimes it's useless, but occasionally it's great. The agentic slot-machine.
So does anyone else ever pull that lever? Drop your longest, most ambitious prompt for me to try out next time I'm going AFK.
A an example, here's one I used recently that actually turned up some useful results:
Take a look at the codebase and the docs in docs/ and wherefore/. Come
up with a plan to systematically evaluate the architecture of the
application from multiple perspectives including, but not limited to:
The overall architecture
The backend architecture
The GraphQL schema
The Postgres database schema
The frontend architecture
The cloud infrastructure (AWS in terraform/)
The frontend tests (e2e, etc.)
The backend tests (integration and unit)
The goal for each is the same: Find ways to simplify, improve, and
streamline the architecture for the following priorities listed from
most important to least: 1. ease of development, extensibility, and
maintainability, 2. robustness and stability, 3. performance. The
end-user functionality must not change or regress. The goal of this
first task is to write up a detailed plan of suggested changes for
each. Present as many potential improvements for each as you feel is
necessary. Rank each potential change on three axes: effort - the
amount of work to rewrite and test the change, risk - how much risk of
breaking the rest of the codebase or system there is, severity/impact -
how critical is it to do and/or how much potential positive impact it
will have.
For each perspective report, rank all possible improvements and present
them from highest priority to lowest. Fan out and send as many as 10
sub-agents out to complete this task.
After all the reports are written, fan out new subagents for each to
review them. The new subagents should antagonistically review each
report to determine how valid it is, whether the suggestions would
cause more problems in the overall system, or other problems, and edit
the reports based on their findings.
Finally, review all of the findings and reports and consolidate them
all into a single writeup containing architecture overhaul plan. The
plan should explain each issue/fix, allong with it's rankings and weave
it all into a step-by-step remediation plan with a checklist for the
items that should be done.
//Comments