Shark Explorer: building a tool for expert human-agent investigations

When I released LeakCanary in 2015, a whole class of Android bugs became visible, but they still weren't easy to fix. Over the next decade, LeakCanary got better UX and better traces, and I spent years teaching engineers how to read them.

Then AI agents arrived: anyone can paste a leak trace into a chat and get back a confident, plausible, wrong root cause. Ten years of effort, handed to a "stochastic parrot". Do experts still matter?

Of course they do! We've been dumping data at agents, hoping for the best, trusting the writeup. Instead, we need tools that enforce the expert's methodology, give humans and agents the same capabilities, and let them show you their progress & reasoning inside the tool itself.

I'll share what I learnt building Shark Explorer, and the design choices behind it. The goal isn't an agent that investigates for you, it's an investigation you can collaborate on, and take further than you could alone.