Agent reliability degrades with trajectory length, but no benchmark isolates it, dev finds

rio_ARC · reddit · 2026-09-04

A developer working on multi-step agent workflows reports that 5–10 step chains look stable, but longer trajectories produce distinct failure modes: unnecessary replanning and repeated tool calls, early small mistakes propagating into later steps, decaying context usefulness, and retries that raise cost without improving outcomes.

They want quantified curves — task success rate, tool-call accuracy, recovery rate, cost per successful task, human intervention — as a function of trajectory length (10 vs 25 vs 50 steps) with model, tools and task distribution held fixed. The open question: is degradation caused by length itself or by state management, memory, retries and orchestration design?

After surveying LangSmith/LangGraph trajectory evaluation, Lyzr's Agent Studio simulations, and platforms like CrewAI and Letta, they found no benchmark that cleanly isolates trajectory length as a variable, and are asking the community for prior experiments.

Related event: Reddit Debates: When Do Long Agent Trajectories Break Down?(2 posts)→

Original post →

More from coding & agent

coding & agent channel →