We sped up bun by 100x

7 hours ago

Install the vers CLI tool to prepare swarm infrastructure and log in.
Set up environment variables like GITHUB_API_KEY, VERS_API_KEY, and ANTHROPIC_API_KEY to authenticate agents.
Write an initial plan in a markdown file and point the coding agent to it, specifying that agents should run in VMs.
The project involved rewriting Git in Zig (ziggit) using agent swarms, focusing on performance, WebAssembly compilation, and Bun integration.
Agents run in infinite loops with scripts that handle git operations (sync, commit, push) and execute coding tasks based on prompts.
The ziggit project achieved significant speedups: up to 85.4x for findCommit on M4 Mac, 34.3x for cloneBare on Linux, and overall 4x to 10x faster than Git.
Ziggit's WebAssembly binary is 5.4x smaller than wasm-git, with 68 named exports for better usability.
Succinct mode is enabled by default to reduce output verbosity, saving tokens for LLM usage.
The theory behind agent swarming emphasizes precise prompts and human oversight, akin to managing a team with clear delegation.
Key lessons include the need for regular check-ins, avoiding over-reliance on agents, and applying Kernighan's Law to maintain debuggability.

Hasty Briefsbeta