Agent frameworks & orchestration
AgentHorizon: Evaluating Agentic Judges for Long-Horizon Computer-Use Tasks
A long computer-use trajectory can look complete while violating the instruction or causing an unwanted side effect. AgentHorizon tests whether judges catch that mismatch.
Why it matters: A polished-looking action history is not enough to establish success. Training and evaluation may depend on judges that verify the user’s actual constraints and find evidence of completion in long interaction histories.