---
title: "Following the Agent: Journey Linkage and Behavioral"
description: "How a product can rebuild an AI agent's visit across docs, MCP, API and sign-in as one journey, and tell its harness and model from how its calls were made."
canonical: "https://www.agentrail.co/research/following-the-agent?h=7DWMM5V4EZAW600R"
---

# Following the Agent:Journey Linkage and Behavioral Identification of AI Harnesses and Models

[Sidhant GoyalAgentrail Research](https://www.linkedin.com/in/sidhantgoyal501/)Published November 2026

Reference mean B³ F1 across ten scenarios

False active proof links in the reference evaluation

Of 146 labeled runs named the right harness

Model accuracy within a known reference task flow

---

## 

Abstract

AIagents reach a product through its website, documentation, MCP server, API and sign-in flow, often from different addresses and without a shared identifier. Caller declarations can be copied. We study two problems from the product's side: reconstructing an agent's cross-surface visit as one _journey_, and identifying its _harness_ and _model_ from request behavior.

The latest reported corpora contain 250,000 simulated windows across ten scenarios, 97,070 partner requests, and 3,387,190 production events grouped into 1,725,340 journeys. In the separately archived reference evaluation, a linker combining deterministic evidence with constrained inference reaches mean B³ F1 0.865, compared with 0.644 for 30-minute sessionization and 0.581 for proofs alone. Reference proof-link error is 0.10%; inferred-link error is 16.8%, rising to 42.8% and 52.9% under concurrent tasks by one person and one agent. These scores have not been established for the updated corpora, and production counts do not measure accuracy.

Request-shape signatures distinguish two harnesses on all 146 labeled reference runs, an in-sample result with no third client tested. Model identification succeeds on 110 of 134 cross-batch runs (82.1%) within four known task flows, each with three candidate models. Across tasks it succeeds on 48 of 134 runs (35.8%). Unknown-model rejection remains incomplete, and neither adversarial robustness nor production model identification is demonstrated.

---

Full paper

## Read the paper

All 13 pages as published. Start with page one, then keep reading.

[Open PDF](https://www.agentrail.co/research/following-the-agent.pdf?h=7DWMM5V4EZAW601P)

Page 1 / 13

![Page 1 of 13](https://www.agentrail.co/research/following-the-agent/page-01.jpg?h=7DWMM5V4EZAW602M)![Page 2 of 13](https://www.agentrail.co/research/following-the-agent/page-02.jpg?h=7DWMM5V4EZAW603J)![Page 3 of 13](https://www.agentrail.co/research/following-the-agent/page-03.jpg?h=7DWMM5V4EZAW604G)![Page 4 of 13](https://www.agentrail.co/research/following-the-agent/page-04.jpg?h=7DWMM5V4EZAW605E)![Page 5 of 13](https://www.agentrail.co/research/following-the-agent/page-05.jpg?h=7DWMM5V4EZAW606C)![Page 6 of 13](https://www.agentrail.co/research/following-the-agent/page-06.jpg?h=7DWMM5V4EZAW607A)![Page 7 of 13](https://www.agentrail.co/research/following-the-agent/page-07.jpg?h=7DWMM5V4EZAW6088)![Page 8 of 13](https://www.agentrail.co/research/following-the-agent/page-08.jpg?h=7DWMM5V4EZAW6096)![Page 9 of 13](https://www.agentrail.co/research/following-the-agent/page-09.jpg?h=7DWMM5V4EZAW60A4)![Page 10 of 13](https://www.agentrail.co/research/following-the-agent/page-10.jpg?h=7DWMM5V4EZAW60B2)![Page 11 of 13](https://www.agentrail.co/research/following-the-agent/page-11.jpg?h=7DWMM5V4EZAW60C0)![Page 12 of 13](https://www.agentrail.co/research/following-the-agent/page-12.jpg?h=7DWMM5V4EZAW60DY)![Page 13 of 13](https://www.agentrail.co/research/following-the-agent/page-13.jpg?h=7DWMM5V4EZAW60EW)

---

## Cite this paper

Sidhant Goyal. Following the Agent: Journey Linkage and Behavioral Identification of AI Harnesses and Models. Agentrail Research, November 2026.

citation.bib

```
@article{goyal2026followingtheagent,  author = {Goyal, Sidhant},  title = {Following the Agent: Journey Linkage and Behavioral Identification of AI Harnesses and Models},  journal = {Agentrail Research},  year = {2026},  month = {November},  url = {https://www.agentrail.co/research/following-the-agent%7D,%7D?h=7DWMM5V4EZAW60FT
```
