Two agents book the same flight. One searches once, compares three fares and books. The other opens the wrong site, retries a failing form six times, and gets lucky. Their final answers match; their trajectories don’t. A trajectory is the trace that exposes the difference: the sequence of thoughts, actions and observations an agent produced on the way to its answer. In LLM tracing tools it usually appears as a tree of spans, one per model call or tool call.
Trajectories matter in three places. In AI evaluation, trajectory-level checks catch wasted steps, unsafe tool calls and right answers reached for the wrong reasons. In debugging, they show where a run went off course. In training, successful trajectories become examples for fine-tuning or reward signals for agentic reinforcement learning. Longer trajectories are harder: every extra step is another chance to fail, and errors compound.
In episode 63 at 24:08, Yutori’s Dhruv Batra says what makes agentic tasks hard is long-horizon trajectories, and that letting a web agent write a compound action in code cuts the trajectory length. In the same episode at 31:11 he describes the agentic trajectories his company’s navigator agents generate as a valuable source of data.
Hear it from the guest
“What makes agentic tasks hard is long horizon trajectories”
Quotes lightly edited to remove filler words.
Sources
- Yao et al.: ReAct, Synergizing Reasoning and Acting in Language Models (arXiv 2210.03629) — The ReAct paper interleaves reasoning traces and actions and analyzes the resulting trajectories on question answering and interactive decision-making tasks; checked October 24, 2026.
- Zhou et al.: WebArena, A Realistic Web Environment for Building Autonomous Agents (arXiv 2307.13854) — Benchmark of long-horizon web tasks in which agents act step by step across realistic websites and are scored on functional task completion; checked October 24, 2026.
