I've been building agents on web search APIs (@p0, @ExaAILabs, @brave) and got curious: when something new is published, how long until search can find it?
So I built a small open experiment to find out. A few findings 🧵
Your AI router made 10,000 decisions yesterday.
Can you defend one of them? Why that model, what it cost, whether the alternative would have held up, who approved the policy?
I built AgentRoute: receipts for every AI routing decision. 🧵
I built a quality score for AI-generated plans. It returned 100/100 for every plan.
Not a model problem. My formula saturated at five steps, and every plan has five steps.
A metric that never disagrees with you is decoration.
A failed agent run shows you what happened. It won't tell you which earlier event to investigate — and a final score can't localize the moment it broke. That gap is where the debugging hours go.
Every coding agent I use is brilliant and completely amnesiac.
It writes code, I correct it, the PR merges, and that conversation is gone. Next session starts from zero.