AI Agents Built This App in 29 Days, 11.5 Billion Tokens, 113,890 Tool Calls. 2,895 Sessions. скачать в хорошем качестве

AI Agents Built This App in 29 Days, 11.5 Billion Tokens, 113,890 Tool Calls. 2,895 Sessions.

Built with GSD — A lightweight and powerful meta-prompting, context engineering and spec-driven development system for Claude Code by TÂCHES. https://github.com/taches-ai/gsd ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ This is a 2-hour cinematic visualization of an entire software project built by Tibsfox and Claude over 25 days (Jan 31 – Feb 28, 2026). Every file creation, every agent spawn, every tool call — captured and rendered into a six-panel real-time dashboard synchronized to music. THE NUMBERS • 11.6 billion tokens processed • 113,890 tool invocations across 26 distinct tools • 2,895 sessions totaling 601 hours of compute time • 256 unique actors: 1 human, 1 Claude main, 1 GSD orchestrator, 253 specialized sub-agents • 104 Executors, 35 Planners, 39 Verifiers, 20 Researchers, 5 Orchestrators • 17,642 gource log events across 25 active development days • 223,256 rendered frames at 30fps • 648 GSD workflow commands coordinating the entire build WHAT YOU'RE SEEING Six synchronized panels, each showing a different dimension of the development process: Panel A — File Tree (top-left, large): A Gource visualization of the living repository. Files colored by domain — pink for agents, blue for core, green for tools, cyan for platform, purple for packs, orange for application, teal for composition. 256 actors appear as avatars: Tibsfox (human developer) with real photo, Claude and GSD with their logos, and hundreds of sub-agents with distinctive shapes — green diamonds for Executors, purple squares for Planners, pink shields for Verifiers, cyan crosshairs for Researchers, orange rings for Orchestrators. Synthetic nodes under /Flow/, /Teams/, and /Comms/ make inter-agent collaboration visible as beams of light between actors. Panel B — Token Flow (top-right): Scrolling flame-style strip showing token consumption — input, output, cache creation and cache reads stacked and color-coded. Reveals the bursty nature of AI development: intense planning followed by sustained execution. Panel C — Complex Plane (top-right corner): A mathematical embedding mapping daily development activity into the complex plane, each day plotted as a point whose position encodes the ratio of activity types. The trajectory traces evolution from early exploration to focused execution. Panel D — Tool Markov Flow (bottom-left): Directed graph of transition probabilities between the top 10 most-used tools. Nodes in a circle, sized by cumulative usage. Edges pulse with color on real-time transitions, revealing Read→Edit→Bash→Grep workflow rhythms. Panel E — Spectrogram (bottom-center): Scrolling audio spectrogram synchronized frame-by-frame with the music. Panel F — Session Timeline (bottom-right): Real-time activity chart showing active sessions, tool calls per minute, and agent spawn events. Color-coded bars distinguish human sessions from autonomous agent work. HOW IT WAS MADE The visualization pipeline was itself built by the same human-AI team it visualizes — a meta-creative loop where Claude Code and GSD built the tools to render their own development story. Raw data was extracted from 2,895 Claude Code session transcripts into structured JSONL files: tool invocations, token usage, session metadata, GSD commands, tool transition probabilities, and daily metrics. An 8-script pipeline transforms this data into the final video: Scripts 00/00b/00c generate the enhanced gource log (merging git history with agent activity and synthetic collaboration nodes), 256 unique avatar images via ImageMagick, and milestone captions from GSD phase transitions. Script 01 renders the gource file tree to raw video — 2560×1620 at 30fps, with seconds-per-day calibrated to match the 7,442-second audio duration. Each of the 25 active days plays out proportionally across the full runtime. Scripts 02/03 precompute panel data (token histograms, transition matrices, complex plane coordinates) and pre-render scrolling strip images for the spectrogram and token flow panels. Script 04 composites all six panels into 1920×1080. Panels rendered per-frame using pycairo and matplotlib, with interval-based caching — complex plane every 10 frames, Markov graph every 3 — achieving 18fps render throughput across all 223,256 frames. Script 05 muxes the synchronized audio with crossfade. Total render time: ~4 hours on a single machine with GPU-accelerated H.264 encoding via NVENC. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 🎵 Music: "November Waves" by Tibsfox   / novemberwaves