zjunlp/onedayagent_traj

OneDayAgent Trajectory Data

Execution trajectories and LLM-as-judge scores for all experiments reported in OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents (Zheng et al., 2026), evaluated on the AgentIF-OneDay benchmark (104 tasks, 767 instance-level rubric points).

This bundle contains the raw evidence chain for every reported run: the full agent trajectory, the final deliverable artifacts, and the per-criterion judge scores. Every headline number in the paper is directly reproducible from the files here, with no external dependencies.

---

1. Runs included

Nine runs, each covering the full 104-task AgentIF-OneDay suite. Six main backends (Section 3.2 / Table 3) plus one additional baseline and three ablation variants (Section 3.3 / Table 4).

Run directory	Paper role	Backend LLM	Paper overall score
onedayagent_glm52_20260623_132100_8210	OneDayAgent (main)	GLM-5.2	0.821
onedayagent_gemini31propreview_20260513_*	backend variant	Gemini-3.1-Pro-Preview	0.743
onedayagent_qwen35_397b_20260610_*	backend variant	Qwen3.5-397B-A17B	0.708
onedayagent_qwen35_9b_20260507_*	backend variant	Qwen3.5-9B	0.624
onedayagent_qwen36_27b_20250618_*	backend variant	Qwen3.6-27B	0.613
codex_gpt55_20260622_174300_6643	baseline	Codex (GPT-5.5 medium)	0.664
ablation_study/ablation_glm52_react_direct_*	DIRECT	GLM-5.2	0.771
ablation_study/ablation_glm52_decompose_only_*	DECOMP	GLM-5.2	0.804
ablation_study/ablation_glm52_verify_only_*	VERIFY	GLM-5.2	0.804

The FULL ablation variant is not duplicated under ablation_study/; it is identical to the main onedayagent_glm52_* run (both modules enabled).

---

2. Uniform per-run layout

Every run has the same six components.

<run>/
├── auto_score_<ts>.jsonl        # 767 lines — per-rubric judge scores
├── auto_score_<ts>.txt          # human-readable aggregate of the above
├── run_<ts>.log                 # runtime log
├── env_snapshot.txt             # runtime environment config
├── <Backend>_<ts>/              # backend subdir, contains only:
│   └── rollout1.jsonl           # 104 lines — one full trajectory per task
└── taskif_<id>_<ts>/            # 104 dirs — final deliverable artifacts per task
    └── ...                      # whatever the agent produced (xlsx/png/md/pptx/...)

These six items form the complete evidence chain (task definition → agent execution → final artifact → judge score) for every run.

---

3. rollout1.jsonl — the core trajectory file

One JSON object per line, 104 lines per run. Same schema for every backend, including the Codex baseline (only trajectory.ext_info.agent differs, ReactAgent vs CodexAgent).

Top-level fields (per line)

Field	Type	Content
question_id	str	e.g. taskif_111
title, description	str	task statement
attachment_filenames	list[str]	user-provided input files
score_criteria	list[obj]	all rubric points: {content, score} (the 767 total)
reference_answer_attachment_filenames	list[str]	reference deliverables
task_tag	str	interaction pattern: Open Workflow Execution / Latent Instruction Inference / Iterative Refinement
domain_tag	str	Work / Life / Study
rubrics_tag	str	Execution / Content / Form
time	str	time budget: <1h / 1-4h / 4-8h / 8-12h / 12-24h / 24+h
question	str	full prompt sent to the agent
prediction	str	agent's final textual answer
time_cost	float	wall-clock latency in seconds (Table 3 Latency column)
result_files	list[str]	final deliverable filenames (match taskif_<id>_* contents)
task_ts	str	per-task start timestamp YYYYMMDD_HHMMSS
trajectory	obj	full conversation (see below)

trajectory sub-object

trajectory:
  guid            : str   — trajectory id
  system_message  : {token_cost, role, content, tool_specs}   — system prompt + tool schemas
  conversations   : list[stage]   — ordered execution stages (see below)
  ext_info        : {type, model, agent, task_description, task_seed}
  create_time     : str   — ISO timestamp

conversations — execution stages

Each run is split into ordered stages. A typical full OneDayAgent task has:

idx	stage	questions	solutions	answer
0	planning	task + planner output	—	subtask JSON list
1..n	subtask	subtask prompt	ReAct turns (reason/act/observe)	subtask summary
n+1	synthesis	synthesis prompt	—	candidate deliverable
n+2	verify	verification prompt	—	{completed, reason, missing_items, suggestions}
n+3	repair?	(only when verify fails) repair feedback + ReAct turns	repaired deliverable

The DIRECT ablation has a single subtask stage and no verify/repair. The DECOMP variant has subtask decomposition but no verify/repair. The VERIFY variant has verify/repair but no decomposition.

---

4. auto_score_*.jsonl / .txt — judge scores

auto_score_*.jsonl — 767 lines (one per rubric criterion, summed across 104 tasks)

{
  "question_id": "taskif_111",
  "agent_name": "react",
  "method": "gemini-3.1-pro-preview",   // the judge model
  "criterion_content": "The returned file accurately names the subtable \"March\"...",
  "criterion_score": 1,                  // 0 or 1
  "satisfied": true,
  "reasoning": "The answer successfully created a new worksheet named 'March'..."
}

method is the LLM-as-judge (Gemini-3.1-Pro-Preview, temperature 0.1, 65 536 max tokens; see Table 2). agent_name is always the literal "react" regardless of the actual backend — identify runs by directory name, not this field.

auto_score_*.txt — pre-aggregated report

Contains the headline number and all Table 3 / Table 4 breakdowns (by task type, domain, rubric dimension, time budget, with/without attachments). The Average score line is exactly the paper's normalized overall score ×100.