Doubao-Seed-Evolving Tested: Coding, Multimodal, and Long-Horizon Agent Skills

Model Upgrade Overview

Doubao-Seed-Evolving is a continuously iterated Seed model version optimized for coding and agent scenarios, integrated into ByteDance's Volcano Agent Plan. The plan provides access to domestic flagship models, Seedance Skills, web search, and memory enhancement. Once connected to a harness, the model always serves the latest capabilities without version management.

Development Task Evaluation

PPT to Travel Agency Website

The author tested the model by converting a PPT into a travel agency website. The model autonomously selected a "luxury adventure magazine" design style, sourced corresponding brand colors and textures, and after development performed self-testing to detect layout issues, iteratively adjusting until optimal.

Personal Project Management Tool

A custom tool combining checklist and short-term calendar for managing N short-cycle tasks with subtasks on specified dates. The model replicated Apple Reminders' interaction in list mode, and supported a "read-and-burn" project management flow where completed projects are deleted.

The author notes the agent capability is strong: each task is seriously decomposed, developed, tested, and fixed before completion.

Multimodal and Visual Capabilities

The model shows improved image content understanding, visual reasoning, and visual risk identification, with enhanced code understanding and generation in multimodal scenarios.

"背影" Reading Comprehension + SVG Generation

First attempt misinterpreted the prompt, exhibiting common quotation issues. After correction, the model successfully designed orange and silhouette elements on the canvas with good aesthetics, and produced copy without inappropriate line breaks.

Simon Willison's Classic Test (SVG Animation)

The author extended the test to generate an animated SVG, which performed well.

3D Rubik's Cube Animation

The model was tasked to create a single HTML file with a canvas animation of a 3D Rubik's Cube using simulated perspective (no WebGL, no libraries). The prompt required specific color orientation, standard notation, and a sequence of animated scrambling and solving with smooth face turns, continuous rotation, step-by-step and automatic solving modes, and mouse drag for view change.

Create a single HTML file with a canvas animation: a 3D Rubik's Cube rendered with
simulated perspective on the 2D canvas (no WebGL, no libraries). Orientation: white on
top, green facing front, red on the right. Use standard notation: R/L/U/D/F/B = clockwise
quarter turn of the right/left/up/down/front/back face (viewed from that face), an
apostrophe = counterclockwise. Sequence: (1) Show the solved cube slowly rotating for
2 seconds. (2) Scramble it with exactly these 10 animated face turns, one at a time:
R, U, F', D, L', B, R', U', F', D'. (3) Pause 2 seconds. (4) Solve it with exactly these
10 animated face turns: D, F, U, R, B', L, D', F, U', R'. (5) End on the solved cube
rotating slowly. Each face turn must be smoothly animated (~0.5s), with correct sticker
colors tracked through every move, visible gaps between stickers, and shading based on
face orientation. The cube keeps slowly rotating in space throughout. The whole sequence
plays automatically without requiring user interaction; additionally, support step-by-step
solving and automatic solving, and allow mouse dragging to change the viewing angle.

Long-Horizon Skills Execution

The author tested the most complex skill, "z-web-pack" (a personal knowledge base builder), which involves over 30 minutes of execution with many steps: pre-check environment, create package directory, page-by-page scraping (GitHub channel → HTTP direct with Referer → Markdown direct save → Jina fallback), expand content links (depth 1-2, keep only documents/papers/reports), image extraction (lazy-load true URL, srcset max resolution, magic number suffix correction, deduplication, filter avatars/badges/small images, pixel classification), video extraction (direct mp4 download, browser recheck on entry page), generate six manifests, and final verification (check residual external images, avatar domains, file completeness).

During execution, after a page scraping failure, the model performed layered troubleshooting, wrote two patch scripts, rebuilt the summary, and conducted three rounds of final verification. The author concludes that if this skill passes, other skills can be considered validated.

Conclusion

In agent and coding scenarios, visual and multimodal understanding is essential: the model must see and understand to screenshot, critique, and self-correct, closing the task loop. Doubao-Seed-Evolving and Doubao Search are available via Volcano Agent Plan, with a limited-time monthly plan starting at 9.9 CNY.