Evaluation Benchmarks for Agents Move Beyond Single-Turn Accuracy Metrics

## Why Old Benchmarks Fall Short A benchmark measuring whether a model gets one question right doesn't capture whether an agent can complete a ten-step task where an error at step three quietly derails everything after it. ## The New Metrics Newer evaluation approaches increasingly measure full task completion rate, recovery from intermediate errors, and efficiency -- how many steps or tool calls a task took -- rather than scoring only the final output in isolation. ## What This Means for Model Selection Teams choosing a model for agent work are increasingly looking past general benchmark leaderboards toward task-completion-specific evaluations that better predict real-world agent reliability.