p-doom/AGI-CAST-0.6k

---

---

AGI-CAST: Behaviour-Cloning Knowledge Work

AGI-CAST-0.6k is a >600-hour dataset of screencasts capturing raw workflows of researchers at p(doom). This is the first major release in our effort to collect months-long fine-grained expert trajectories of human reasoning in their day-to-day work.

Please refer to the blog post for more details.

Dataset Summary

Contemporary frontier model development has saturated the internet and increasingly relies on manual data labeling (including handcrafting RL environments) to elicit the next set of model capabilities. We work towards methods of training on the billions of hours of uncaptured screencasts produced by the digital work force. While crowd-code was our first effort to capture real-world long-horizon research engineering workflows by recording fine-grained IDE interactions, AGI-CAST goes one step further. AGI-CAST-0.6k provides 646 hours of raw screencasts of p(doom) members (Franz, Mihir, Alfred, and Yll) conducting AGI research. The dataset captures project inception, system design, literature review, debugging cycles, training run monitoring, eval creation, IDE and browser usage and paper exploration, among many other things.

Key Features

• Total Duration: 646.21 hours
• Expert Trajectories: Raw videos of real-world work (not synthetic nor curated demonstrations)
• Long-Horizon: Single sessions up to 12 hours, collected by recording p(doom) members over the course of three months

Dataset Statistics

The dataset consists of 147 individual sessions. The dataset is organized in subfolders corresponding to each contributor.

Folder	Videos	Avg Duration	Median Duration	Total Hours	Frame rate	Resolution
Total	147	4:23:46	3:18:21	646.21
alfred	18	3:21:20	2:22:33	60.40	1280x720	30.000
franz	82	4:33:21	3:52:36	373.57	1280x720	30.000
mihir	40	4:27:46	3:13:37	178.51	1280x720	30.000
ylli	7	4:49:06	4:36:04	33.73	1920x1080	60.000

Attribution

We thank our colleagues (Mihir, Alfred, Yll) for contributing their work sessions to the dataset. If you use AGI-CAST in your research, please cite our work:

@article{srambical2025agi-cast:,
  author = {Srambical, Franz},
  title = {AGI-CAST: Making Agents Work Like Humans},
  journal = {p(doom) blog},
  year = {2025},
  note = {https://pdoom.org/blog.html}
}