An AI agent fails halfway through a job. What should it remember when it tries again?
That question led me to build AgentTrace, a Python and FalkorDB project for keeping track of agent runs. It records task progress and decisions in a graph so you can inspect what happened and see where a run stopped. That’s especially useful when several agents share a workflow.
When you debug an agent run, what do you look for first: the failed step, the input that caused it, or the decision that sent it down that path?
AI made video production cheap. It didn't make it easy.
The tools are everywhere — generate a shot, clone a voice, cut a trailer, grade a frame. Anyone can press the button. Almost no one gets a result worth watching.
Because the tool was never the bottleneck. The bottleneck is knowing what to make: the shot that carries meaning, the cut that lands on the beat, the three seconds that were supposed to make you feel something. AI can produce ten thousand variations. It cannot tell you which one is true.
That is not a limit of the model. That is the job.
The people who win with this don't hand the craft to the machine. They keep the intent, the taste, the decision — and let the machine carry the labour.
The multiplication only works if both sides are strong.
Weak operator, strong AI → fast garbage.
Strong operator, weak AI → slow, expensive, ordinary.
Strong operator, strong AI → the shot that used to be impossible.