Abdu07/agentglass-swerebench-human-verified-annotations

AgentGlass: human-verified first-breakdown steps in coding-agent trajectories

500 failed OpenHands agent trajectories, each read by hand and annotated with the first step at which the run became unrecoverable. 484 of them carry a verified step, a root-cause category and one recommended change to the agent.

The trajectories come from nebius/SWE-rebench-openhands-trajectories and are not redistributed here. This file joins to them on trajectory_id.

What this is for

Outcome benchmarks tell you whether an agent failed. They do not tell you where the run went wrong, or why. Every row here is a run the benchmark's own hidden tests marked as failed, with a human-located breakdown step, the reasoning behind it and a verbatim quote from the trace supporting it.

Every row is an oracle-confirmed failure. Nothing here rests on a model's opinion of whether a run succeeded.

Contents

500 rows, 26 columns, one row per reviewed trajectory.

	Rows
reviewed by hand	500
with a verified breakdown step	484
marked unlocatable	16
revisited in a second pass	398
with a root-cause category	484
with a recommended fix	484
with an evidence quote	500

The written content is 72,412 words of justification and 51,889 words of adjudication, all of it human.

The 500 were drawn from the 1,521 oracle-confirmed failures in a run over 3,000 trajectories. Of the 500, 416 agents claimed success in their final message and 84 gave up without calling finish().

The human columns and the model columns are kept separate on purpose. Drop every model column and what remains is still a usable annotation set.

Columns

Identity and provenance

Column	Source	Meaning
trajectory_id	upstream	join key, the chatcmpl-… id used by the Nebius dataset
dataset	this work	always nebius

Oracle

Column	Source	Meaning
oracle_succeeded	upstream	always 0 here: only failures were reviewed. The only correctness signal used anywhere in this work
oracle_source	this work	where the label came from, metadata.resolved

Agent self-report

Column	Source	Meaning
l1_agent_claimed_success	derived	1 if the agent's final message claims success, empty if it never called finish()
l1_agent_finish_message	upstream	the agent's final message, verbatim. The only upstream text reproduced in this file

Automatic localisation

Column	Source	Meaning
l2_model	this work	the model that produced the proposal, gpt-5.4-nano
l2_coherence_score	model	the judge's own 0 to 1 score. Kept because it is part of what the layer returned, but it carries little signal and should not be used as a quality measure
l2_first_breakdown_step	model	the proposed step index, empty on the rows where the judge abstained
l2_breakdown_reason	model	one sentence from the judge

The human review

Column	Source	Meaning
reviewed	this work	True on every row of this file
l2_human_step	human	the verified first-breakdown step. This is the label everything downstream uses
l2_human_verified	human	1 where a step was located, 0 on the 16 unlocatable rows
review_verdict	human	confirm (39), correct (445) or unlocatable (16)
review_confidence	human	the reviewer's confidence, 0 to 1
review_reasoning	human	the written justification, citing step numbers
review_evidence_quote	human	a verbatim quote from the trajectory supporting the step
adjudicated	this work	True on the 398 rows revisited in a second pass
adjudication_reasoning	human	the second-pass note

Root cause, on the 484 verified rows

Column	Source	Meaning
l3_failure_category	model	one of the ten AgentRx labels
l3_failure_category_id	model	its numeric id
l3_critical_step_index	this work	the verified step, written back so it cannot drift from l2_human_step
l3_confidence	this work	a provenance flag, not a model confidence: 0.90 marks a real judge output
l3_explanation	model	the judge's reasoning at that step

Recommended fix, on the 484 verified rows

Column	Source	Meaning
l4_fix_type	model	prompt, tool, logic or environment. Collapsed to logic on 481 of 484 rows and should not be relied on
l4_recommended_fix	model	one change to the agent, 2 to 5 sentences. A change to the agent's prompt, tools, policy or environment, not a patch to the repository it was editing

How the review was done

The 500 were drawn by sorting the 1,521 oracle-confirmed failures by identifier and taking from the top, which is as good as random under the assumption that the ids are independent of what happened in the run. That assumption holds here: the sample's step and coherence distributions match the population's. The protocol for each trajectory was the same:

1. read the issue text and ask what a correct fix would have to do;
2. read the trajectory from step 0, in the normalised thought / action / result form, asking at each step whether a competent agent could still have recovered after it, and point at the first step where the answer is no;
3. record that step, a verbatim evidence quote from the trajectory, and a confidence between 0 and 1, or record unlocatable if no agent step causes the failure;
4. record the verdict against the automatic proposal and write a justification citing step numbers.

The reviewer points at their own step. It agrees with the automatic proposal on some trajectories and not on others, and either way it is the step that counts: root-cause attribution receives it, and never receives the automatic one.

Reviewer confidence

Recorded on all 500 rows. Mean 0.626, median 0.60, range 0.35 to 0.90.

Band	Rows
0.35 to 0.49	54
0.50 to 0.64	224
0.65 to 0.79	160
0.80 to 0.90	62

Confidence is a reviewer judgement about how cleanly a trajectory isolates a single step, not a probability. Filter on it if you want a stricter subset.

What the annotations show

• Failures are committed mid-trajectory. The median verified breakdown step is 32, and 94% of verified failures fall at step 15 or later.
• The automatic localiser finds the region but not the step. Its exact proposal survived review on 8.4% of the 462 rows where it named a step and the reviewer located one, correlating with the verified step only weakly (Spearman rho 0.39) and missing by a mean of 13.5 steps on both sides.
• Root causes over the 484: Plan Adherence Failure 283, Misinterpretation of Tool Output 77, Invalid Invocation 44, Invention of New Information 33, Intent-Plan Misalignment 21, System Failure 13, Under-specified User Intent 9, Inconclusive 4.

Using it with the trajectories

import pandas as pd
from datasets import load_dataset

ann  = pd.read_csv("agentglass_final_dataset.csv")
traj = load_dataset("nebius/SWE-rebench-openhands-trajectories", split="train").to_pandas()

df = ann.merge(traj, on="trajectory_id", how="inner")
verified = df[df.l2_human_verified == 1]

Limitations

• One annotator, so there is no inter-annotator agreement figure.
• One agent and one task family: OpenHands on SWE-rebench. The failure profile may not carry to other scaffolds or other kinds of task.
• None of the 484 recommended fixes has been applied and re-run, so their usefulness is argued from their specificity rather than measured.
• l4_fix_type collapsed to logic on 481 of 484 rows. It is published as it came out rather than quietly dropped, and it should not be used.
• l3_confidence is a provenance flag, not a model confidence.
• unlocatable is a reviewer judgement about what was discoverable from inside the workspace. It is defensible row by row, but it is not a benchmark-level defect count.

Human authorship

Every value in l2_human_step, l2_human_verified, review_verdict, review_confidence, review_reasoning, review_evidence_quote and adjudication_reasoning was produced by a person reading a trajectory. No model wrote, drafted or suggested any of them.

Licence and attribution

Released under CC-BY-4.0. The upstream trajectories are nebius/SWE-rebench-openhands-trajectories, also CC-BY-4.0, collected with Qwen3-Coder-480B-A35B-Instruct under OpenHands v0.54.0 on tasks from nebius/SWE-rebench. Please credit Nebius alongside this dataset.

Citation

@mastersthesis{elmoustapha2026agentglass,
  author = {El Moustapha, Abdellahi},
  title  = {AgentGlass: An Asymmetric Glass-Box Evaluation Framework for
            Autonomous Agent Trajectories},
  school = {aivancity School for Technology, Business and Society},
  year   = {2026}
}