[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-agent-trajectory::en":3,"gloss-cluster-agent-trajectory::en":26,"gloss-next-agent-trajectory::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"agent-trajectory","agents","Agent Trajectory","A trajectory is the full recorded path an agent took through a task: every model call, every tool invocation with its arguments, every result returned, and the final answer, in order. It is the agent equivalent of a request trace, and it is the unit you evaluate, debug and replay against. The reason it matters is that a final answer alone is not enough evidence to judge an agent. Two runs can produce the same correct output while one called three tools and the other called eleven, retried a failing endpoint, and got there by luck; only one of those is a system you would put in front of customers. Conversely a wrong answer says nothing about where the failure was — bad retrieval, a mis-parsed tool result, a plan that never recovered from an early error — and without the trajectory the team is left guessing. Practical trajectory work has three parts. Capture: log the whole sequence including arguments and raw tool outputs, with the timing and token cost of each step. Evaluation: score the path as well as the destination, since step count, tool-selection accuracy and recovery behaviour are the things that predict cost and reliability in production. And replay: keep enough of the record that a stored trajectory can be re-run against a changed prompt or a new model, which is what turns a pile of past incidents into a regression suite.","An agent trajectory is the ordered record of every model call, tool use and result in a run — the unit you evaluate, debug and replay agents against.",null,[11,14,17,20,23],{"slug":12,"name":13},"agent","Agent",{"slug":15,"name":16},"agent-loop","Agent Loop",{"slug":18,"name":19},"distributed-tracing","Distributed Tracing",{"slug":21,"name":22},"eval-harness","Eval Harness",{"slug":24,"name":25},"llm-as-judge","LLM-as-Judge",[27,31,35,38,39,42,45,48,51,54,57,61],{"slug":28,"category":5,"name":29,"updated_at":30},"agent-budget","Agent Budget","2026-08-24T02:46:37+00:00",{"slug":32,"category":5,"name":33,"updated_at":34},"agent-checkpointing","Agent Checkpointing","2026-08-24T02:46:38+00:00",{"slug":36,"category":5,"name":37,"updated_at":30},"agent-handoff","Agent Handoff",{"slug":15,"category":5,"name":16,"updated_at":30},{"slug":40,"category":5,"name":41,"updated_at":30},"agent-memory","Agent Memory",{"slug":43,"category":5,"name":44,"updated_at":34},"agent-sandbox","Agent Sandbox",{"slug":46,"category":5,"name":47,"updated_at":30},"agentic-rag","Agentic RAG",{"slug":49,"category":5,"name":50,"updated_at":34},"computer-use","Computer Use",{"slug":52,"category":5,"name":53,"updated_at":30},"context-engineering","Context Engineering",{"slug":55,"category":5,"name":56,"updated_at":34},"dry-run","Dry Run",{"slug":58,"category":5,"name":59,"updated_at":60},"escalation-path","Escalation Path","2026-08-24T03:30:02+00:00",{"slug":62,"category":5,"name":63,"updated_at":30},"human-on-the-loop","Human-on-the-Loop"]