AI Trajectory Tracing

 Kaushik Rangarajan, Senior Architect, Wipro Limited

Introduction

In a rapid evolving landscape of 2026, the “black box” problem of artificial intelligence is no longer being met with a shrug. As we move from simple chatbots to autonomous agents that manage our finances, health and legal workflows, the demand for transparency has birthed a critical new discipline: “AI Trajectory Tracing.”

If machine learning was once about the destination (the output), Trajectory Tracing is about the odyssey (the process).

https://hackmd.io/@alexaa34/B18x_lUnWl

https://medium.com/@alexharris59600/ai-trajectory-tracing-bc55bfd56ba6

From Outputs to Pathways: What is Trajectory Tracing

Historically we evaluated AI using “black box” metrics like Accuracy or F1-Score. If the AI gave the right answer, it was good. But in an agentic world, being right for the wrong reasons is a liability. AI Trajectory Tracing is the systematic logging and analysis of AI’s internal reasoning, tool interactions, and environmental feedback over time. Instead of looking at a single point of time, we look at the entire “flight path” the model took to reach its conclusion.

A standard AI Trajectory consists of four distinct layers:

1. The Thought (Reasoning): The internal “Chain of Thought” where the model plans its next move.

2. The Action (Tool Use): The specific external calls made (ex: searching a database, executing Python Code).

3. The Observation (Feedback): The Raw data the environment sends back to the AI.

4. The Stata (Context): The evolving memory of the agent as it moves through a multi-step task.

Why 2026 is the Year of the “Glass Box”

Three reasons are behind the move toward trajectory-aware evaluation:

1) High-Score Illusion: We have learned the hard way that a model can hit a high score on a benchmark by “hallucinating” its way to the right answer or exploiting data leaks. Trajectory tracing exposes these circuitous or unsound paths. A “High Utility” score now requires that every step in the reasoning chain be grounded in evidence.

2) Mechanistic Interpretability: As organisations push towards “AI Lie Detectors,” trajectory tracing provides the data. By tracing “circuits” within the model during a specific trajectory, researchers can identify if a model is being deceptive — internally planning one thing while outputting another.

3) Regulatory Demand: With the EU AI Act entering its full application phase in August 2026, “explainability” is no longer a feature — it is legal requirement for high-risk systems. Trajectory logs serve as the black box flight recorder for legal audits.

Technical Implementations Traces and Spans

In practise, trajectory tracing borrows heavily from distributed software tracing:

1) Trace: The entire end-to-end journey of a user request

2) Span: A Single Unit of work within the journey (ex: one vector database lookup)

Modern observability stacks now allow developers to stich these spans together into a visual timeline. If an agent fails, you do not just see “Error 500”, you see that the agent failed at Step 4 because it misinterpreted the schema of a SQL Database.