A developer with zero video codec experience used Claude Code agent teams to build Sinter, a working video codec from scratch. The goal: test whether one-shot agent workflows could tackle video compression. The codec works. Files are 18.6 times larger than H.264 at comparable quality.
About 5,000 lines of C across 12 improvement loops, according to the project's GitHub repo. The simulated team included personas based on real experts like Google's Jim Bankoski and Mozilla's Timothy Terriberry. Sinter uses lapped transforms, perceptual vector quantization, and rANS entropy coding as a patent-free alternative to the H.264 lineage. Perceptual quality holds up. Compression doesn't.
The gap comes from missing tools that H.264 built up over decades. No sub-pixel motion compensation. No B-frames. Fewer entropy contexts than CABAC. Each missing piece multiplies the size penalty. The project estimates a realistic patent-safe ceiling at 4-6x H.264 size. Matching H.264 would mean rebuilding its entire toolset.
For agent watchers, this is a useful benchmark. Someone with no domain knowledge shipped functional software in a field where specs run 500+ pages. But it also shows where one-shot approaches hit walls. Video compression gains come from accumulating dozens of interlocking optimizations over years. Agents got the architecture right. The details are where H.264 wins.