trajectory

in English: agent trajectory

advancedspreading

A trajectory is the complete recorded path of an AI agent during a task: its written-out reasoning, steps, tool usage, and the results it received, in sequence, from start to finish.

In plain terms

When you entrust an AI agent with a task, it does not give a single answer but carries out a series of steps: it considers the situation, searches, opens a file, checks the result, then decides how to continue. The recorded sequence of this entire path is the trajectory.

It is like a fixed route in a hiking app. You see not only that you reached the summit but also where you went, where you got lost, where you turned back, and how long it took.

This matters because the final result alone says little. Two agents can reach the same good outcome, with one going straight to it while the other circled needlessly, ran up costs, or did something it should not have. You can only see this from the path.

This differs from a simple chat history: the trajectory includes not only spoken words but also actions and their consequences.

How it works

The agent receives the task and the starting situation: this is the first entry of the trajectory.
The model decides what the next step should be, such as searching, opening a file, or running a program; this is often accompanied by a brief explanation.
The step is executed, and its result (search results, error message, file content) is returned to the agent as an observation.
This decision–action–observation loop repeats until the agent deems the task complete, gets stuck, or reaches the set limit.
The system logs the entire sequence, so it can be replayed and evaluated later, used for debugging, or even for training new models.

An everyday example

You ask an AI assistant to find weekend accommodation in Eger, and a collapsible row appears above the response: '12 steps: search, 3 pages opened, prices compared'. If you expand it, you see a simplified, summarized view of the trajectory—what it did step-by-step before drafting its offer. The full, detailed log is typically visible only to developers.

Who is affected

developerThe trajectory is the raw material for debugging and evaluation: it reveals at which step the agent went off track. → Code

entrepreneur, executiveIf you entrust work to an agent, the trajectory shows exactly what it did, how much it cost, and whether it worked reliably.

employeeBefore accepting an agent's work, you can review its steps to verify whether it reached the result from a reliable source and via a sound path. → Work & HR

lawyer, citizenIn case of dispute or damage, the recorded sequence of steps can serve as important evidence of what the agent did and on what instructions, provided the log was kept complete and authentic. → Law

A common misconception

MythIf the agent delivers a good result, it does not matter how it got there.

In realityA good final result may be due to luck, unnecessary cost, or a step that should not have been taken. Therefore, agents are increasingly evaluated based on their path, not just their outcome.

Limits and risks

The trajectory can be long and hard to read: a complex task may consist of hundreds of steps, which a human rarely reviews in full.
An agent's stated reasoning does not necessarily reflect exactly what the model actually based its decision on, so it is worth observing its actions and their results.
Sensitive data (password, customer data, details of internal documents) can end up in the log, so its storage and sharing must be guarded just like the data itself.
The same task may run via a different path the second time, so general reliability of the agent cannot be inferred from a single trajectory.

Similar, but not the same

chain of thought – Chain of thought is just the model's step-by-step reasoning; trajectory additionally includes the actual actions taken and their results.
context window – The context window is what the model sees at once; the trajectory is the entire path taken, part of which may be pushed out of this window during long tasks.
conversation history – The history is the message exchange between the user and the model, while the trajectory also records the agent's independent steps, tool calls, and observations.

Related concepts

Broader

AI-agent, observability, reinforcement learning

Narrower

tool calling, step, monitoring

On the edge

agentciklus, chain of thought, tracing, agent evaluation, context window

A brief history

The term originally comes from physics (the trajectory of a moving body). In AI terminology, it originates from reinforcement learning, where for decades it has denoted the sequential series of states, actions, and rewards. Its widespread application to language model-based agents began following the 2022 ReAct paper, which described agent operation as a sequence of thought-action-observation steps.

Where things stand

By the end of 2026, it is everyday jargon for agent developers and evaluation teams, and most developer tools record and display it. The average user will mostly encounter it only as a collapsible 'steps' view and rarely hear the word itself.

Status: 2026 November

In the news

Appears in 38 Agent Age stories; the curve shows weekly occurrence in the news magazine.

AI agents are taking over HR processes, but who governs them?Nov 4, 2026 · Work & HR
Mayo Clinic Platform launches EHR-integrated clinical decision support engine across 35+ subspecialtiesNov 3, 2026 · Health
Fear, lack of training and limited decision-making power slow workplace AI adoption in HungaryNov 2, 2026 · Work & HR
Harvey makes acquisition, announces partnership and launches new features for law firmsNov 2, 2026 · Jog
Malaysia trains 329,414 teachers in AI ahead of school curriculum rolloutOct 29, 2026 · Education & Family
Walmart to let outside AI agents shop on its site from next month, with agent profiles and a gatewayOct 21, 2026 · Marketing
NHS England expands Copilot to 500,000 staff as NHS Wales launches training campaignOct 21, 2026 · Health
Siemens and P&G scale AI-based quality inspection globallyOct 19, 2026 · Ipar

Also known as

trajectory · agent trajectory · agent trajectory · trajectory · agent trajectory · run log · step sequence · rollout

Sources

Yao et al.: ReAct: Synergizing Reasoning and Acting in Language Models (arXiv, 2022)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.