Learning part 1: Building AI agents feels like the early 2000s web

Building AI agents right now feels a lot like building a website in the early 2000s. Everyone is doing it, a lot of it is experimental, and somewhere in that pile is the stuff that quietly changes how things work.

Right now I think the interesting use is agents that do useful, boring, time costly things. The errands that are genuinely worth doing but not worth your hour. The chasing. The comparing. The “is this actually a good deal or does it just look like one.” Basically, outsource the legwork and some of the judgement. Not all of it.

This month’s attempt is a points hunter called 🦆 DuckFleet. The premise sounds like a joke: in 2025, rubber ducks at Big W could be stacked across promos into enough Qantas points for a business-class seat. Real deal, real maths, but catching it meant refreshing forums A LOT. So I built a fleet of agents to do that hunting overnight and just tell me what is actually worth doing.

It is open source, and I am building it in public: the wins, the failure modes, and the parts that are honestly still way too technical for a normal person to use (there are plenty). It is also my entry for Google’s All Things Agentic hackathon.

So which of your useful but time-costly errands would you actually hand to an agent, and which would you never trust it with?