A team of researchers has released an open-source tool that lets you watch an AI coding agent's reasoning unfold step by step, instead of just checking whether it got the right answer.
Graphectory Viewer is a web-based tool built on the team's earlier Graphectory representation, which converts raw agent logs - thoughts, actions, and observations - into structured graphs. The tool works across multiple agent frameworks, so it is not locked to one vendor's format. Users can inspect individual runs node by node, search and filter across large collections of trajectories, and view Sankey-style diagrams showing how an agent moves between problem-solving phases. The team released the tool alongside documentation, precomputed graphs, and a large trajectory corpus so others can reproduce the work.
This matters because most agent evaluation still comes down to pass or fail on a benchmark, which says nothing about why an agent succeeded or where it wandered off track. A viewer that exposes the actual decision path lets researchers compare a working run against a failed one and pinpoint the exact phase where things went sideways - useful for anyone debugging agent behavior rather than just scoring it.
It is a research artifact, not a polished product. But process-level tooling like this tends to get absorbed into commercial agent-debugging platforms once the underlying idea proves its worth.