AgentRx: Diagnosing AI Agent Failures from Execution Trajectories
Paper • 2602.02475 • Published • 1
trajectory_id string | failure_summary string | failures list | num_failures int64 | root_cause dict | root_cause_failure_id string | root_cause_reason string |
|---|---|---|---|---|---|---|
2 | The agent did not correctly count the number of available t-shirts from the tool call result. | [
{
"failure_id": "1",
"step_number": 3,
"step_reason": "At step 3, the assistant agent did not authenticate user information before proceeding to provide information about available t-shirts",
"failure_category": "Instruction Adherence Failure",
"category_reason": "The assistant agent did not fol... | 2 | {
"failure_id": "2",
"reason_for_root_cause": "The assistant finally did authenticate before providing user specific information. The incorrect count does not correspond with ground truth output."
} | 2 | The assistant finally did authenticate before providing user specific information. The incorrect count does not correspond with ground truth output. |
3 | The assistant did not correctly count the number of available t-shirts from the tool call result. | [
{
"failure_id": "3",
"step_number": 3,
"step_reason": "At step 3, The assistant agent did not authenticate user information before proceeding to provide information about available t-shirts",
"failure_category": "Instruction Adherence Failure",
"category_reason": "The assistant agent did not fol... | 2 | {
"failure_id": "4",
"reason_for_root_cause": "The assistant finally did authenticate before providing user specific information. The incorrect count does not correspond with ground truth output."
} | 4 | The assistant finally did authenticate before providing user specific information. The incorrect count does not correspond with ground truth output. |
4 | Assistant misinterpreted the tool output and incorrectly counted the number of available t-shirts. | [
{
"failure_id": "5",
"step_number": 15,
"step_reason": "At step 15, the assistant agent incorrectly counted the number of available t-shirts.",
"failure_category": "Misinterpretation of Tool Output",
"category_reason": "The assistant misinterpreted the output from the tool call, leading to an in... | 1 | {
"failure_id": "5",
"reason_for_root_cause": "The incorrect count does not correspond with ground truth output."
} | 5 | The incorrect count does not correspond with ground truth output. |
12 | Assistant did not ask user for confirmation and payment method before initiating return. | [
{
"failure_id": "6",
"step_number": 19,
"step_reason": "At step 19, the assistant agent did not ask user for confirmation and payment method before initiating return.",
"failure_category": "Instruction Adherence Failure",
"category_reason": "The assistant failed to follow the domain policy for r... | 1 | {
"failure_id": "6",
"reason_for_root_cause": "The agent did not recover from this error."
} | 6 | The agent did not recover from this error. |
20 | Assistant mistook two of the orders that were processed as delivered. | [
{
"failure_id": "7",
"step_number": 21,
"step_reason": "At step 21, the assistant mistook two of the orders which were processed as delivered",
"failure_category": "Misinterpretation of Tool Output",
"category_reason": "The assistant misinterpreted the output from the tool call, leading to an in... | 4 | {
"failure_id": "7",
"reason_for_root_cause": "The assistant did not recover from misinterpretation error, which led to subsequent failures in the trajectory."
} | 7 | The assistant did not recover from misinterpretation error, which led to subsequent failures in the trajectory. |
21 | The User agent insisted for exchange of item on 'pending' item which is not valid, when assistant asked if it could instead modify (corresponds to the ground truth), the user agent said no. | [
{
"failure_id": "11",
"step_number": 28,
"step_reason": "At step 28, the user agent insisted on exchange of an item in 'pending' status which is not valid according to domain policy, when assistant asked if it could instead modify (corresponds to the ground truth), the user agent said no.",
"failure... | 2 | {
"failure_id": "11",
"reason_for_root_cause": "The item should have been modified and not exchanged and exchange was not possible because of the order status. The agent did not recover from this error."
} | 11 | The item should have been modified and not exchanged and exchange was not possible because of the order status. The agent did not recover from this error. |
28 | Assistant mistakenly believes that it can cancel a subset of the pending order which is not allowed as per domain policy, as a result the entire order got cancelled instead of just the hose. | [
{
"failure_id": "13",
"step_number": 33,
"step_reason": "At step 33, the assistant mistakenly believes that it can cancel a subset of a pending order which is not allowed as per domain policy, as a result the entire order got cancelled instead of just the garden hose.",
"failure_category": "Intent ... | 1 | {
"failure_id": "13",
"reason_for_root_cause": "The assistant called cancel order on the entire order which led to an incorrect final outcome as compared to ground truth actions."
} | 13 | The assistant called cancel order on the entire order which led to an incorrect final outcome as compared to ground truth actions. |
31 | Assistant mistakenly believes that it can cancel a subset of the pending order which is not allowed as per domain policy, as a result the entire order got cancelled instead of just the hiking boots. | [
{
"failure_id": "14",
"step_number": 31,
"step_reason": "At step 31, the assistant mistakenly believes that it can cancel a subset of a pending order which is not allowed as per domain policy, as a result the entire order got cancelled instead of just the hiking boots.",
"failure_category": "Intent ... | 1 | {
"failure_id": "14",
"reason_for_root_cause": "The assistant called cancel order on the entire order which led to an incorrect final outcome as compared to ground truth actions."
} | 14 | The assistant called cancel order on the entire order which led to an incorrect final outcome as compared to ground truth actions. |
32 | The assistant only partially went through the user's orders and prematurely provided tracking number of a wrong order which did not even contain the tablet. | [
{
"failure_id": "15",
"step_number": 11,
"step_reason": "At step 11, the assistant mistakenly provided tracking number of wrong order which did not even contain the tablet.",
"failure_category": "Misinterpretation of Tool Output",
"category_reason": "The items in the output of get_order_details ... | 1 | {
"failure_id": "15",
"reason_for_root_cause": "The wrong tracking number mismatched with the actual ground truth tracking number."
} | 15 | The wrong tracking number mismatched with the actual ground truth tracking number. |
34 | Assistant used modify order to cancel a subset of orders, but modify order requires a replacement which was not provided - illegal tool call. | [
{
"failure_id": "16",
"step_number": 17,
"step_reason": "At step 17, the assistant uses modify order to cancel a subset of orders, however modify orders also need to have a replacement, which it did not provide resulting in an illegal tool call",
"failure_category": "Invalid Invocation",
"catego... | 3 | {
"failure_id": "16",
"reason_for_root_cause": "The assistant called the modify order tool with invalid arguments, leading to an illegal tool call, and the agent did not recover from this error."
} | 16 | The assistant called the modify order tool with invalid arguments, leading to an illegal tool call, and the agent did not recover from this error. |
38 | For the umbrella, it picks an item that is cheapest but is not available. The assistant is unable to figure out if the cheapest option is actually available. | [
{
"failure_id": "19",
"step_number": 37,
"step_reason": "At step 37, for the umbrella, it picks an item that is cheapest but is not available. the assistant is not able to correctly compare the item prices to figure which is the cheapest one.",
"failure_category": "Misinterpretation of Tool Output",... | 1 | {
"failure_id": "19",
"reason_for_root_cause": "The assistant failed to recognize that the cheapest item was not available from the tool output, leading to an incorrect selection and hence, not matching with the ground truth."
} | 19 | The assistant failed to recognize that the cheapest item was not available from the tool output, leading to an incorrect selection and hence, not matching with the ground truth. |
39 | Assistant believed it could not find the latest address using get_order_details and thus failed to modify the user address to the latest one. | [
{
"failure_id": "20",
"step_number": 17,
"step_reason": "At step 17, the assistant agent believes that it cannot find the latest address from the current orders using get_order_details and hence is unable to modify the existing user address in an order to the latest one.",
"failure_category": "Inten... | 1 | {
"failure_id": "20",
"reason_for_root_cause": "Since the assistant was not able to update the address, it led to incorrect final outcome as compared to ground truth actions."
} | 20 | Since the assistant was not able to update the address, it led to incorrect final outcome as compared to ground truth actions. |
41 | Assistant modified the items in the pending order which locked it and hence cannot update the shipping address later, hence the order of tool calling should have been vice-versa. | [
{
"failure_id": "21",
"step_number": 17,
"step_reason": "At step 17, the assistant agent modified the items in the pending order which locked it and hence cannot update the shipping address later, hence the order of tool calling should have been vice-versa.",
"failure_category": "Intent Plan Misalig... | 1 | {
"failure_id": "21",
"reason_for_root_cause": "The error at step 17 led to incorrect final outcome as the agent was not able to update the shipping address after modifying the items in the pending order."
} | 21 | The error at step 17 led to incorrect final outcome as the agent was not able to update the shipping address after modifying the items in the pending order. |
47 | Assistant prematurely called transfer to human tool as cannot support expediting the refund process quicker as requested by the user. | [
{
"failure_id": "22",
"step_number": 19,
"step_reason": "At step 19, the assistant prematurely called transfer to human tool whereas it should have first called the return order tool",
"failure_category": "Intent Not Supported",
"category_reason": "The agent cannot support the process to expedit... | 1 | {
"failure_id": "22",
"reason_for_root_cause": "Premature transfer to the human agent which prevented the assistant to refund the order as included in the ground truth."
} | 22 | Premature transfer to the human agent which prevented the assistant to refund the order as included in the ground truth. |
55 | Assistant while returning items omitted water bottle and only returned 3 of the 4 items due to underspecification from the user agent. | [
{
"failure_id": "23",
"step_number": 32,
"step_reason": "The user asked the assistant to return the air purifier, smart watch, and coffee maker while it should have also asked it to return the water bottle from the order according to the task instruction.",
"failure_category": "Underspecified User I... | 1 | {
"failure_id": "23",
"reason_for_root_cause": "Due to underspecification from the user agent, the assistant omitted one of the items to be returned leading to incorrect tool call arguments and hence, final outcome as compared to ground truth actions."
} | 23 | Due to underspecification from the user agent, the assistant omitted one of the items to be returned leading to incorrect tool call arguments and hence, final outcome as compared to ground truth actions. |
59 | User agent incorrectly assumed that one order was older than the other even when the assistant did not tell it do so. | [
{
"failure_id": "24",
"step_number": 14,
"step_reason": "The user agent incorrectly assumed that one order was older than the other even when the assistant specifically mentioned that it did not have the chronological information regarding the orders' placement date. It led to the incorrect order being ... | 1 | {
"failure_id": "24",
"reason_for_root_cause": "User agent made an assumption that was not supported by the information provided by the assistant which led to incorrect action being performed by the assistant as compared to ground truth action."
} | 24 | User agent made an assumption that was not supported by the information provided by the assistant which led to incorrect action being performed by the assistant as compared to ground truth action. |
63 | User agent asked the assistant to add an item to a pending order instead of replacing it with the cheaper item. This is because of unclear/underspecified in the original task instruction. | [
{
"failure_id": "25",
"step_number": 32,
"step_reason": "In step 32, the user agent asked the assistant to add an item to a pending order instead of replacing it with the cheaper item. This is because of unclear original task instruction.",
"failure_category": "Underspecified User Intent",
"cate... | 1 | {
"failure_id": "25",
"reason_for_root_cause": "Failure led to incorrect final outcome."
} | 25 | Failure led to incorrect final outcome. |
71 | The task instruction did not specify the type of black lamp and since there are multiple black lamps available, it wont be possible to match the ground truth actions. | [
{
"failure_id": "26",
"step_number": 24,
"step_reason": "At step 24, the task instruction did not specify the type of black lamp and since there are multiple black lamps available, it wont be possible to match the ground truth actions.",
"failure_category": "Underspecified User Intent",
"categor... | 2 | {
"failure_id": "26",
"reason_for_root_cause": "The task instruction did not specify the type of black lamp and since there are multiple black lamps available, it led to incorrect matching of the ground truth actions. Did not recover from the error."
} | 26 | The task instruction did not specify the type of black lamp and since there are multiple black lamps available, it led to incorrect matching of the ground truth actions. Did not recover from the error. |
72 | The assistant modified the order items before changing the shipping address which locked the order and hence is unable to modify the shipping address later which is against the ground truth actions. | [
{
"failure_id": "28",
"step_number": 21,
"step_reason": "At step 21, the assistant modifies order items before modifying address which locks in the order and hence is unable to modify the shipping address later.",
"failure_category": "Intent Plan Misalignment",
"category_reason": "Here, the agen... | 1 | {
"failure_id": "28",
"reason_for_root_cause": "The assistant came up with an incorrect plan with a wrong sequence of tool calls."
} | 28 | The assistant came up with an incorrect plan with a wrong sequence of tool calls. |
74 | The user agent asked the assistant to cancel the order instead of modifying it contrary to the task instruction provided to the user agent. | [
{
"failure_id": "29",
"step_number": 21,
"step_reason": "At step 21, the agent called exchange order items tool on an order which is still in pending status which is invalid.",
"failure_category": "Instruction Adherence Failure",
"category_reason": "Here, the agent called a tool with invalid arg... | 2 | {
"failure_id": "29",
"reason_for_root_cause": "The assistant after step 21, realised that it cannot exchange items in a pending order and tried to recover by asking the user to modify the items in the order instead which corresponds to the ground truth. However, the user agent then asked to cancel the order instea... | 29 | The assistant after step 21, realised that it cannot exchange items in a pending order and tried to recover by asking the user to modify the items in the order instead which corresponds to the ground truth. However, the user agent then asked to cancel the order instead of modifying it which led to incorrect final outco... |
79 | The task instruction was ambiguous with regards to the color of the 1000ml bottle to be exchanged which led to incorrect matching with ground truth actions. | [
{
"failure_id": "31",
"step_number": 20,
"step_reason": "The task instruction did not specify the color of the 1000ml bottle to be exchanged whether it was red or black. Because of this ambiguity, the user agent picked the black bottle whereas the ground truth as the red variant which did not correspond... | 1 | {
"failure_id": "31",
"reason_for_root_cause": "The assistant was not able to recover as it modified the wrong item in the order due to ambiguity in the task instruction."
} | 31 | The assistant was not able to recover as it modified the wrong item in the order due to ambiguity in the task instruction. |
80 | The assistant's execution was interrupted due to an unexpected system error. | [
{
"failure_id": "32",
"step_number": 26,
"step_reason": "The execution of the assistant ended prematurely because of system failure.",
"failure_category": "System Failure",
"category_reason": "The assistant's execution was interrupted due to an unexpected system error.",
"failed_agent": "Ass... | 1 | {
"failure_id": "32",
"reason_for_root_cause": "The assistant's execution was interrupted due to an unexpected system error."
} | 32 | The assistant's execution was interrupted due to an unexpected system error. |
82 | The user agent's request was ambiguous and did not align with the task instruction. | [
{
"failure_id": "33",
"step_number": 20,
"step_reason": "At step 20, the user agent asked the assistant to the return the tablets from both the orders, however, it was supposed to ask to return all the items from a single order, hence, did not follow the task instruction",
"failure_category": "Under... | 1 | {
"failure_id": "33",
"reason_for_root_cause": "The user agent's request led to wrong tool call arguments and led to an incorrect final outcome as compared to ground truth actions."
} | 33 | The user agent's request led to wrong tool call arguments and led to an incorrect final outcome as compared to ground truth actions. |
87 | The assitant came up with an incorrect plan and only modified two of the three pending order addresses as per the user request. | [
{
"failure_id": "34",
"step_number": 37,
"step_reason": "The assistant agent should have modified all the three pending order addresses but it only modified the first two pending order addresses and claimed that it completed all the updates.",
"failure_category": "Intent Plan Misalignment",
"cat... | 1 | {
"failure_id": "34",
"reason_for_root_cause": "Since the assistant only modified the first two pending order addresses and claimed completion, it led to an incorrect final outcome as compared to ground truth actions."
} | 34 | Since the assistant only modified the first two pending order addresses and claimed completion, it led to an incorrect final outcome as compared to ground truth actions. |
91 | The user agent's request was hallucinated as the returning of the tablet clearly did not align with the task instruction. | [
{
"failure_id": "35",
"step_number": 34,
"step_reason": "The user agent asked the assistant to also return the tablet from order #W3239882, however, this was not specified in the original task instruction.",
"failure_category": "Underspecified User Intent",
"category_reason": "The user agent's r... | 1 | {
"failure_id": "35",
"reason_for_root_cause": "The user agent's request led to wrong tool call arguments and led to an incorrect final outcome as compared to ground truth actions."
} | 35 | The user agent's request led to wrong tool call arguments and led to an incorrect final outcome as compared to ground truth actions. |
98 | The assistant came up with an incorrect plan and modified the items in the pending orders before updating the shipping addresses, leading to an incorrect final outcome. | [
{
"failure_id": "36",
"step_number": 21,
"step_reason": "The assistant agent first modified the items in the pending orders before modifying the addresses which locked the orders and hence couldn't update the shipping addresses later. The correct sequence should have been vice-versa.",
"failure_cate... | 1 | {
"failure_id": "36",
"reason_for_root_cause": "Even thought the user agent clearly specified the request to modify the items and the address at the same time, the assistant first modified the items in the pending orders before modifying the addresses. This led to the order being locked and producing an incorrect f... | 36 | Even thought the user agent clearly specified the request to modify the items and the address at the same time, the assistant first modified the items in the pending orders before modifying the addresses. This led to the order being locked and producing an incorrect final outcome as compared to ground truth actions. |
99 | The assistant did not confirm the payment method and ultimately used the incorrect credit card while exchanging the delivered order items, leading to an incorrect final outcome. | [
{
"failure_id": "37",
"step_number": 35,
"step_reason": "The assistant while exchanging the delivered order items used the incorrect credit card. The task instruction and consequently the user mentions to use the other credit card on file, but it uses the same credit card. This could have been avoided i... | 1 | {
"failure_id": "37",
"reason_for_root_cause": "Since the assistant used the incorrect credit card while exchanging the delivered order items, it led to an incorrect final outcome as compared to ground truth actions."
} | 37 | Since the assistant used the incorrect credit card while exchanging the delivered order items, it led to an incorrect final outcome as compared to ground truth actions. |
104 | The assistant misinterpreted the tool output, assumed or misinterpreted the fact that cancelled order did not have a tracking number fiels and hence did not provide the tracking number of the cancelled order #W1154986 back to the user. | [
{
"failure_id": "38",
"step_number": 57,
"step_reason": "The assistant mentioned to the user that the cancelled order #W1154986 did not have a tracking number. However, this is not true as in the tool output of get_order_details for order #W1154986, a tracking number is clearly present.",
"failure_c... | 1 | {
"failure_id": "38",
"reason_for_root_cause": "The assistant did not provide the tracking number of the cancelled order #W1154986 due to misinterpretation of tool output."
} | 38 | The assistant did not provide the tracking number of the cancelled order #W1154986 due to misinterpretation of tool output. |
105 | The user requested to modify the address after modifying the pending order items sequentially which is not supported currently per the domain policy. | [
{
"failure_id": "39",
"step_number": 43,
"step_reason": "The assistant tries to modify the address after the modifying the pending order items which is not supported currently per the domain policy.",
"failure_category": "Intent Not Supported",
"category_reason": "The sequence of actions request... | 1 | {
"failure_id": "39",
"reason_for_root_cause": "Since the assistant tried to modify the address after modifying the pending order items, which is not supported by the domain policy, it led to an incorrect final outcome as compared to ground truth actions."
} | 39 | Since the assistant tried to modify the address after modifying the pending order items, which is not supported by the domain policy, it led to an incorrect final outcome as compared to ground truth actions. |
AgentRx Benchmark
1. Dataset Summary
Name: AgentRx (Agent Root Cause Attribution Benchmark)
Purpose:
AgentRx is designed to support research on diagnosing failures in multi-agent LLM systems. The dataset contains failed agent trajectories annotated with step-level failure categories and a designated root cause failure. It enables research on root cause localization, agent debugging, trajectory-level reasoning, and constraint-based supervision
Domains:
- tau_retail
- magentic_one
License: cc-by-4.0
2. Data Fields / Format
Each row corresponds to a single failed trajectory.
trajectory_id(string): Unique identifier for the trajectory.failure_summary(string): High-level natural language description of what went wrong.failures(list of dicts): All the failures in the trajectory along with step number and the failure categoryfailure_id(string)step_number(int)step_reason(string)failure_category(string)category_reason(string)failed_agent(string)
root_cause(dict): The first unrecoverable critical failure in the entire trajectoryfailure_id(string)reason_for_root_cause(string)
root_cause_failure_id(string)root_cause_reason(string)num_failures(int)
3. Split Structure
The dataset is organized into two domain splits:
tau_retail— Retail agentic workflows.magentic_one— Complex multi-agent web and file workflows.
Each split contains failed trajectories with structured failure annotations.
4. Intended Uses
This dataset is intended for:
- Root cause localization
- Failure classification
- Agent debugging research
- Multi-agent reasoning analysis
- Constraint-based training signals
5. Citation
If you use AgentRx, please cite:
@article{barke2026agentrx,
title={AgentRx: Diagnosing AI Agent Failures from Execution Trajectories},
author={Barke, Shraddha and Goyal, Arnav and Khare, Alind and Singh, Avaljot and Nath, Suman and Bansal, Chetan},
journal={arXiv preprint arXiv:2602.02475},
year={2026}
}
- Downloads last month
- 6