---
title: "Agent research you can run · Agentrail"
description: "Agentrail research on identifying AI agents, and the harness and model behind each one, from how they behave. Read the paper, then request early access."
canonical: "https://www.agentrail.co/research?h=D43ZT8A392DK800R"
---

# Agent research you can run built for developers.

We're open-sourcing the benchmarks and model weights soon. Until then, request early access and run the same system on your own agent traffic.

[Browse research](https://www.agentrail.co/research?h=D43ZT8A392DK800R#papers)

Algorithm

## Every agent leaves a fingerprint .

Agentrail reads how an agent behaves, not what it claims to be. Mixed, unlabeled traffic goes in; one clean session per agent comes out.

Fingerprint

Hover a point on the print

01Incoming callsunlabeled

- GET`/docs/quickstart`?C
- POST`/v1/search`?G
- CALL`search_docs`?C
- GET`/login`?D
- GET`/login`?D
- GET`/login`?D

03Agents identified

- Claude Code2 callsAnthropic ClaudeReading your docs
- Codex1 callOpenAI GPTCalling your API
- Claude Code3 callsDeepSeekStuck at login

No headers, API keys or user-agent strings needed. The **timing, order and retries** of an agent’s calls form a pattern as unique as a fingerprint, and Agentrail groups every call under the agent that made it.

Journey reconstruction accuracy

Fewer errors than IP or user-agent matching

Hand-offs reconnectedLegacy methods: 0%

Precision on proven links

[Read the methodology](https://www.agentrail.co/research/following-the-agent?h=D43ZT8A392DK801P#sec-5)[Read the paper](https://www.agentrail.co/research/following-the-agent?h=D43ZT8A392DK801P)

Features

## Built on the research. Shipped as tools you can use.

Everything behind the fingerprint, packaged so you can drop it into any stack and see your agents in minutes.

### Behavioural fingerprinting

Unlabeled calls go in on the left; a stable signature per agent comes out on the right. No headers, keys or user-agents needed.

### Session reconstruction

Hand-offs between tools and surfaces are stitched back into one journey per agent.

### Every surface

Web, docs, API and MCP traffic, read through one model.

### Stuck-agent alerts

Know when an agent loops, stalls or gives up, before your users tell you.

```
$ pip install agentrail
from agentrail import trace
trace.agents(app)
```

### Open SDK

A few lines to instrument any stack. MIT licensed.

Research

## Read the paper. Then check our work.

Each paper reports what works, where it fails and how we measured it, with a one-line BibTeX. Benchmarks and model weights are on their way to open source.

### [Following the Agent: Journey Linkage and Behavioral Identification of AI Harnesses and Models](https://www.agentrail.co/research/following-the-agent?h=D43ZT8A392DK801P)

[Sidhant Goyal](https://www.linkedin.com/in/sidhantgoyal501/)

AI agents reach a product through its website, docs, MCP server, API and sign-in flow, with no identifier shared across them. We rebuild each agent’s visit as one journey and identify the harness and model behind it from how its calls were made.

[Read the paper](https://www.agentrail.co/research/following-the-agent?h=D43ZT8A392DK801P)

Install

## One command. Agents, identified.

Install the CLI, point it at your app, and watch every agent show up in your terminal with its own session.

$`pip install agentrail`

Python 3.10+, Node 18+, macOS, Linux and Windows
