Agent frameworks & orchestration
LoopsBench: From Harness Engineering to Loop Engineering in Coding Agent Evaluation
LOOPSBENCH evaluates coding agents on sustained development by making task dependencies, incremental testing, and regression obligations part of the benchmark.
Why it matters: For evaluating agents used in long-horizon software work, this design can reveal whether they follow prerequisites and preserve completed work—not just whether they reach an end state. The reported results also expose gaps in plans and regression behavior.