Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
9
9
task
stringlengths
12
153
messages
stringlengths
2.9k
16.4k
tools_used
listlengths
2
6
num_assistant_turns
int64
2
14
num_tool_calls
int64
2
14
num_messages
int64
7
33
was_resumed
bool
2 classes
traj_0000
Open Firefox, navigate to archive.org, take a screenshot showing the homepage.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_click_element", "computer_type_text", "computer_type_keys" ]
4
4
11
true
traj_0001
press link to youtube, inside youtube search bar, search Numb press enter
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_detect_elements", "computer_click_element", "computer_type_keys", "computer_type_text" ]
8
8
18
false
traj_0002
continue writing on the joke you see on screen
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_detect_elements", "computer_click_mouse", "computer_type_text" ]
7
7
17
false
traj_0003
check weather today in Sofia, having the temperature and ity displayed marks the task with sucess
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_application", "computer_detect_elements" ]
5
5
13
false
traj_0004
Open Firefox, take a screenshot, identify and click the browser's refresh/reload button.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_click_element" ]
2
2
7
true
traj_0005
Take a screenshot. Click on element number 1. Then take another screenshot and describe what changed.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_click_element" ]
3
3
11
true
traj_0006
Open the terminal and run: date > ~/timestamp.txt && cat ~/timestamp.txt. Report what timestamp was written.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys", "computer_screenshot" ]
4
4
11
true
traj_0007
Open Firefox, navigate to example.com, take a screenshot showing the page loaded, then write the page title to /tmp/example_title.txt using the terminal.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_click_element", "computer_type_text", "computer_type_keys" ]
9
9
19
true
traj_0008
Open Firefox, navigate to google.com, click on the search box, type "open source AI", press Enter, and take a screenshot of the results.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_screenshot", "computer_click_element", "computer_type_text" ]
4
4
12
true
traj_0009
Open the terminal and run: echo "test" | sha256sum. Report the hash value.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_screenshot", "computer_type_text" ]
4
4
11
true
traj_0010
Open Firefox, go to example.com, then use Ctrl+L to focus the address bar, type https://example.org and press Enter. Take a screenshot.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_keys", "computer_type_text" ]
4
4
9
true
traj_0011
open the firefox, After its open, search the weather in Sofia
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_application", "computer_type_keys", "computer_type_text" ]
7
7
15
false
traj_0012
close firefox, that will be a completed task
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_type_keys" ]
3
3
8
false
traj_0013
close everythin, when you see desktop again its a completed task mark
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_type_keys" ]
5
5
12
false
traj_0014
check top news for anthropic today
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_screenshot", "computer_type_keys", "computer_type_text", "computer_wait" ]
8
7
16
false
traj_0015
check space free opening the terminal and eecuting command: df -h
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_screenshot", "computer_type_text", "computer_type_keys" ]
4
4
9
false
traj_0016
inside terminal that is open, create a file and write inside some jokes
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_type_text", "computer_paste_text", "computer_type_keys" ]
9
9
13
false
traj_0017
create a file, called jokes.txt inside of it write a joje
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_application", "computer_type_text" ]
9
9
17
false
traj_0018
close the terminal, that will be a complte task
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_type_keys" ]
3
3
8
false
traj_0019
open Firefox and search for AI news"
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_click_element", "computer_type_text", "computer_type_keys" ]
4
4
10
true
traj_0020
open Firefox and search for AI news
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_click_element", "computer_type_text", "computer_type_keys" ]
4
4
10
true
traj_0021
Open the terminal and type: echo "Agent test OK" then press Enter
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys" ]
4
3
10
true
traj_0022
Open Firefox browser
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_click_element", "computer_screenshot" ]
4
4
12
true
traj_0023
Open the terminal and create a file called test_result.txt with the content "phase1 test passed"
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys" ]
8
8
14
true
traj_0024
Open the terminal, check what directory you are in with pwd, then list the files with ls
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys" ]
4
4
9
true
traj_0025
Open VS Code
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_click_element" ]
4
2
11
true
traj_0026
Take a screenshot. Then click on element number 0 that you see in the numbered screenshot
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_screenshot", "computer_click_element" ]
3
3
11
true
traj_0027
Open Firefox, then take a screenshot and click on the address bar (it should be a numbered element near the top)
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_click_element", "computer_screenshot" ]
4
4
12
true
traj_0028
Open the terminal, run the command: date > ~/timestamp.txt && cat ~/timestamp.txt
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys" ]
5
5
10
true
traj_0029
open a terminal, run the command: echo 'hello world', and tell me the output
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys" ]
3
3
7
true
traj_0030
Open the terminal and type: echo "Agent training ok" then press Enter. Report what the terminal shows.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys" ]
3
3
7
true
traj_0031
Open the terminal and create a file called /tmp/training_01.txt with the content "trajectory data collection started".
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys" ]
4
4
9
true
traj_0032
Open the terminal and type: echo hello world then press Enter. Report what you see.
[{"role": "system", "content": "You are a desktop automation agent operating a Linux (XFCE) workstation.\n\nAvailable applications: Firefox, VS Code, Terminal, File Manager.\n\nWorkflow: Observe (screenshot) \u2192 Plan \u2192 Act \u2192 Verify.\n- Always take a screenshot before and after actions.\n- Screenshots show ...
[ "computer_application", "computer_type_text", "computer_type_keys", "computer_screenshot" ]
4
4
11
true
End of preview. Expand in Data Studio

Desktop Agent Trajectories — Sample

AI desktop agent execution trajectories recorded on a Linux (XFCE) workstation. Each trajectory is a complete conversation between an AI agent and a desktop environment, including tool calls (click, type, scroll, screenshot, etc.) and their results.

What's in this dataset

  • 56 unique trajectories of an AI agent completing real desktop tasks
  • Tasks: web navigation, terminal commands, file operations, UI interaction
  • Full conversation in OpenAI chat format (system + user + assistant + tool roles)
  • Tool calls: computer_click_element, computer_type_text, computer_screenshot, computer_application, etc.
  • Screenshots stripped to reduce size — text descriptions preserved

Schema

Field Type Description
id string Trajectory identifier
task string The user task given to the agent
messages string (JSON) Full conversation in OpenAI chat format
tools_used list[string] Unique tools called in this trajectory
num_assistant_turns int Number of agent reasoning/action turns
num_tool_calls int Total tool invocations
num_messages int Total messages in conversation
was_resumed bool Whether the task was interrupted and resumed

Use cases

  • Fine-tuning computer-use agents (convert messages JSON to your training format)
  • Benchmarking agent tool-use accuracy and efficiency
  • Research on GUI grounding and desktop automation
  • Few-shot prompting for agent systems

Format

Each row's messages field is a JSON string in OpenAI chat completion format with tool_calls and tool roles. Parse with:

import json
from datasets import load_dataset

ds = load_dataset("diogoneno/desktop-agent-trajectories-sample")
messages = json.loads(ds["train"][0]["messages"])

Feedback

This is a free sample for community testing. We'd love to hear:

  1. Is this format useful for your agent training pipeline?
  2. What metadata would make this more valuable?
  3. Would you use a larger curated dataset (500+ scored trajectories)?

Open a discussion on this repo — all feedback shapes the full release.

License

Apache 2.0

Downloads last month
29