Wild-Trajectory

27,803 agent trajectories collected in the wild: 80 models, 50 harnesses, 12 task types, 19 public sources. Every trajectory is one Harbor ATIF-v1.8 document, and every document validates against harbor==0.23.0.

This is the dataset of the paper On the Behavioral Traits of LLM Agents (code: github.com/JoeZhao527/A-B-D). Companion datasets: A-B-D/wild-trajectory-annotations (functional behavior annotations) and A-B-D/wild-trajectory-liwc (linguistic behavior, LIWC-22).

Format

One row per trajectory. The columns are the top-level fields of its ATIF document:

column	type	content
schema_version	string	ATIF-v1.8
session_id	string	trajectory id
trajectory_id	string	trajectory id, <source>/<run>/<task>
agent	struct	name (harness id), version, model_name (canonical model id), extra (JSON string)
steps	JSON string	the ATIF step list: one source=agent step per model call, with reasoning_content, message, tool_calls and the observation it received; user and system steps in between
final_metrics	JSON string	totals (steps, tokens, cost where the source reports them)
extra	JSON string	everything beyond the ATIF vocabulary, under extra["wild"] (see below)

Parse the JSON columns to get the full ATIF document back:

import json
from datasets import load_dataset

ds = load_dataset("A-B-D/wild-trajectory", split="train")
row = ds[0]
doc = {**row,
       "agent": {**row["agent"], "extra": json.loads(row["agent"]["extra"])},
       "steps": json.loads(row["steps"]),
       "final_metrics": json.loads(row["final_metrics"]),
       "extra": json.loads(row["extra"])}

wild = doc["extra"]["wild"]
print(doc["agent"]["model_name"], doc["agent"]["name"], wild["task"]["task_type"])

extra["wild"]

ATIF allows custom fields only under extra, so everything this corpus adds lives under extra["wild"] at the level it belongs to.

level	fields
trajectory (extra.wild)	source (slug, license, generator_class, env_realism, selection_channel, fidelity); task (benchmark, native_task_id, task_type, task_subtype, instruction, environment, ground_truth); outcome (score_norm ∈ [0, 1], score_type, score_raw, exit_status); trial; metrics; flags (reasoning_tier, parse_tier, has_user_sim, …); participants; extra (source-specific)
agent (agent.extra.wild)	model_raw, harness_raw (the upstream strings), model_id, harness_id
step (steps[i].extra.wild)	kind (agent, user_msg, system_prompt, env_event, score_event, lifecycle, …); scripted (user text delivered by the runner); is_demo (few-shot steps); summary_only / redacted (reasoning that is an API summary or hidden); harness_injected; call
tool call / observation result	raw_args, args_not_object, call_id_synthesized; modality (image observations are placeholders, pixels are not included), error, content_full (size and hash of truncated content)

Field reference for the usual grouping variables:

variable	location
model	agent.model_name
harness	agent.name
source	extra.wild.source.slug
task type	extra.wild.task.task_type
outcome	extra.wild.outcome.score_norm

How it was built

1. Collect raw trajectories from public releases (19 sources, each frozen at a pinned upstream revision).
2. Parse each source's native log format into a common event stream, and resolve model, harness and task strings to canonical identifiers.
3. Render every trajectory as an ATIF-v1.8 document and validate it with Harbor.
4. Keep trajectories that can be placed in a model × harness × task design, are linear (no branching), and have ≥ 5 annotatable agent steps.
5. Form (model, task type, harness) cells with ≥ 10 trajectories each, and keep models whose cells span ≥ 2 task types, so the subset forms a connected core for cross-context analysis.
6. Draw at most 30 trajectories per cell, balancing successes and failures.

Composition

task type	trajectories
swe	5,332
science	4,689
terminal	3,028
ml-engineering	2,664
security	2,661
tool-api	2,100
research-qa	2,089
other	1,913
dialogue	1,694
web	1,113
office	340
gui	180

source	trajectories	source	trajectories
terminalbench-community	11,254	multi-swe-bench	420
hal	4,876	theagentcompany	330
terminal-bench-leaderboard	3,042	tau2-bench	300
exgentic-traces-v2	2,114	naturebench-traces	282
liveclawbench	2,009	openhands-eval-outputs	240
malt	893	osworld2	180
apiflow-bench	570	gaia-traces	90
toolathlon	570	cybench	60
mcpmark	510	lhtb-leaderboard	33
		tau-bench	30

License

The trajectories come from public releases, each under its own license; see extra.wild.source.license and the original releases before reuse.

Citation

@article{zhao2026behavioral,
  title   = {On the Behavioral Traits of LLM Agents},
  author  = {Zhao, Haokai and Gao, Jie and Xiao, Yunze and Wang, Xintao and Xuan, Weihao and
             Joshi, Aditya and Dredze, Mark and Huang, Jen-tse},
  journal = {arXiv preprint arXiv:2609.32776},
  year    = {2026}
}