owenisas/video-agent-episodes

Video Agent Episodes

Complete, unedited recordings of an AI coding agent (Claude Opus 5.5) making short original videos from written briefs with HyperFrames (HTML, CSS, GSAP, Canvas and WebGL rendered frame by frame to MP4), each graded by a vision judge (Claude Fable 5.1) under a published rubric.

Every episode keeps every model message, every tool call with its exact input and output, every image the agent looked at, the agent's project, the delivered video and the judge's full answer. It is meant for studying and training agents that make visual media: their planning, self-review from rendered frames, and failure modes.

Episodes	502 (181 briefs; samples r2: 24, s1: 181, s2: 111, s3: 52, s4: 83, s5: 51)
Videos delivered	429 (131.2 min of footage)
Labels	PASS 131, PASS_WITH_NITS 204, FAIL 152, unjudged 15
Positive-eligible (pass, no major defect, no failed gate)	214
Agent turns / tool calls / images viewed	23,180 / 12,405 / 3,445
hyperframes render / snapshot calls	502 / 1,494
Actor tokens (output; cache read)	18,097,463; 706,446,930
Aspect ratios	16:9 236, 9:16 180, 1:1 86

Last refreshed 2026-10-27 06:21 UTC. The table is regenerated from episodes.jsonl.

Files

• episodes.jsonl: one row per episode (the table the viewer shows). Key columns: id, brief_id, sample, title, family, genre_id, concept, aspect, duration_s, difficulty, label, host_rule, host_gates_failed, positive_eligible, six score_* columns (0–10), defects_critical/major/minor, turns, tool_calls, images_read, renders, snapshots, token counts, video, archive, archive_sha256, and brief (the full brief as a JSON string).
• videos/<id>.mp4: the delivered video, silent, H.264.
• episodes/<date>/<id>.tar.zst: the full episode. episodes/<date>/<id>.json: its manifest (brief, verdict, usage, per-file sha256, excluded files).
• meta/: the brief bank (brief-bank-v1.jsonl) and its JSON schema, the judge rubric (rubric-v1.2.json with its write-up, earlier versions in rubric-archive/), and the actor prompt (actor-prompt.md).

Inside an archive (zstd -dc <id>.tar.zst | tar -xf -):

Path	Content
brief.json	The brief exactly as the agent received it (including the rubric's banned defaults)
messages.jsonl	Every Agent SDK message in order: assistant text, thinking blocks, tool use, tool results, system and result messages
tool-events.jsonl	PreToolUse / PostToolUse hook records: exact tool input, full output, duration
observations/	Copies of every image the agent opened with Read (what it actually saw)
transcript/	The Claude Code session JSONL, including its request log: the system prompt and tool definitions sent (api-request-shape), per-request parameters (api-request) and message bodies (api-request-blob)
work/	The HyperFrames project the agent built (vendored JS runtimes removed, see Licensing)
out.mp4	The deliverable
frames/	The frames the judge received (JPEG) and their labels
judge.json	The judge request metadata, its full parsed answer, mechanical facts, host verdict and gates
result.json	Run status, turns, tool counts and token usage

How episodes were made

• Actor: Claude Opus 5.5 on Amazon Bedrock (effort high), driven by the Claude Agent SDK for Python (0.2.174), which runs the Claude Code CLI (2.1.310) headless with its built-in tools (Bash, Read, Write, Edit, ...). System prompt: the Claude Code preset plus meta/actor-prompt.md and the brief. Time limit 1,500–3,120 s per brief, turn cap scaled with it.
• Sandbox: one Docker container per episode (Node 22, HyperFrames 0.8.203, headless Chrome with software WebGL, ffmpeg; 4 CPUs, 8 GB). No internet: the model API is the only reachable host. JS libraries (GSAP, three.js, matter-js, p5, d3, lottie-web) are preinstalled locally; only system and HyperFrames-bundled fonts.
• Briefs: 181 briefs across 15 families. Bank v1 (meta/brief-bank-v1.jsonl, 111 briefs) came from a multi-agent research pass over video genres and formats (YouTube, Shorts, explainers, UI, data visualization, physics, typography, loops, title sequences and more) with critic rounds and name clearance. Bank v2 (meta/brief-bank-v2.jsonl) adds 3D-heavy and new-style genres (dioramas, walkthroughs, toon characters, rigid bodies, cloth, particles, attractors, 3D data, mechanisms, extruded type, raymarching, pixel art, ASCII, risograph, comics, paper cut-outs); its briefs are schema-checked but had no critic round, and their coined names are pending (not web-cleared). Every product, place and person named on screen is invented. Several independent samples per brief (sample s1-s5; rN = rerun after a harness fix).
• One-shot prompts (task_type = one-shot): short prompts people posted publicly to test models (Reddit, X, Hacker News, blogs), adapted to a video deliverable and stripped of real brands and characters. The agent sees only one_shot_prompt plus the format; source in one_shot_source_url, the verbatim original in the brief's one_shot.quote, and the survey in meta/one-shot-prompts.md.
• Judge: Claude Fable 5.1 (Claude Opus 5.5 when Fable was rate-limited; judge_model per row) sees 14 frames chosen by the rubric's sampler (first and last frame, success-criteria times, beat midpoints, named events, even fill) plus mechanical facts measured on the video (black and freeze ranges, perceptual still checks, loop seam, render budget, whether the agent looked at its final render). The host recomputes the verdict from the judge's counts and scores (rules R1–R8) and applies host gates; any failed gate makes the label FAIL.
• Re-hosting: when a host-side check was corrected (for example the loop-seam gate on 2026-10-27), labels were recomputed from the stored judge answer without a new model call (rehosted_at and previous_final_label in judge.json). Early batches were graded under rubric v1.0, later ones under v1.2 (rubric_sha256).

Families

Family	Episodes
ui-product	53
explainer	47
scene-3d	47
brand-broadcast	42
data-viz	40
character	38
physics-sim	36
youtube-segment	36
story-medium	34
typography	27
generative-art	27
maps-time-scale	22
music-rhythm	19
social-format	19
loop-ambient	15

Intended uses and limitations

Uses: analysis of agent behaviour on creative coding tasks; supervised fine-tuning or preference data for video-making agents (positive_eligible marks the cleanest positives, FAIL episodes are negatives with the judge's stated reasons); evaluation of judges and rubrics.

Limitations:

• One actor model and one judge model; the judge is itself a model and makes mistakes (it is strict on exact on-screen wording, and sees 14 frames, not every frame).
• Reasoning is not included: Claude Code requests adaptive thinking with display: "updates", so thinking blocks arrive with an opaque signature and almost never with text.
• Videos are silent (the bank has no sound briefs yet).
• renders counts Bash calls that run hyperframes render directly; renders inside scripts the agent wrote are not counted, so it is a lower bound.
• Samples of one brief share the prompt and are not independent tasks.
• For episodes judged before 2026-10-27 02:00 UTC whose brief has a hold requirement, the black/freeze facts were not passed to the judge (a host bug since fixed); labels were not affected.
• Costs and token counts are as reported by the SDK and Bedrock; no prices are implied.
• harness_error = true (85 rows): the episode died because its Bedrock key was revoked mid-run (HTTP 403), not because of the agent; exclude these from any analysis or training. The v2 and one-shot briefs that died on the first key were re-run as <id>-r2; most of those re-runs died on the second key.
• label = null (15 rows): never judged. Model access ended before the judge reached them; the episode itself (video, recordings) is complete.

Licensing

The dataset (briefs, rubric, prompts, recordings, judge outputs, projects and videos) is released under CC BY 4.0. Content was generated with Anthropic's Claude models; make sure your use also complies with Anthropic's Usage Policy and terms.

One-shot prompt texts (one_shot.quote, and the adapted one_shot.prompt) remain their authors'; they are quoted with attribution for research and are not covered by the CC BY license.

The Claude Code system prompt and tool descriptions recorded in transcript/ (and the SDK's init messages) are Anthropic's text, included as they were sent; they are not covered by the CC BY license.

Third-party components referenced by the projects:

• HyperFrames (Apache-2.0): the CLI that renders the projects; its documentation appears in tool outputs where the agent read it.
• Fonts used in renders: Inter, EB Garamond, JetBrains Mono (bundled with HyperFrames), Noto, Liberation (SIL OFL 1.1) and DejaVu (Bitstream Vera / public domain). Font files are not included.
• JS runtimes the projects load from assets/vendor/ are not included in the archives (identical in every episode); vendor_files in each manifest lists their names and sha256: GSAP 3.14.2 (GSAP Standard "no charge" license), three.js 0.186.1 (MIT), matter-js 0.20.0 (MIT), p5.js 2.3.4 (LGPL-2.1), d3 7.9.0 (ISC), lottie-web 5.13.0 (MIT). Install the same versions to re-render a project.

Privacy and safety

No personal data and no third-party media: episodes ran without internet access, and every name is invented. Archives were scanned for credential-shaped strings before release; redacted_keys in a manifest counts replacements.

Loading

from datasets import load_dataset
from huggingface_hub import hf_hub_download

ds = load_dataset("owenisas/video-agent-episodes", split="train")
row = ds[0]
video = hf_hub_download("owenisas/video-agent-episodes", row["video"], repo_type="dataset")
archive = hf_hub_download("owenisas/video-agent-episodes", row["archive"], repo_type="dataset")
# zstd -dc <archive> | tar -xf -

Citation

@misc{video_agent_episodes_2026,
  title  = {Video Agent Episodes: Claude Opus 5.5 building motion-graphics videos with HyperFrames},
  author = {owenisas},
  year   = {2026},
  url    = {https://huggingface.co/datasets/owenisas/video-agent-episodes}
}