Zerostack – A Unix-inspired coding agent written in pure Rust

Original Article ↗ Hacker News Discussion ↗

Performance & Resource Usage

  • Zerostack reports ~8 MB RAM idle and ~11–12 MB with a 128k context loaded; startup claimed around 90 ms.
  • Many contrast this with Claude Code, Pi, and Opencode, which are reported to use multiple GB of RAM, become sluggish, leak memory, and even trigger OOM on modest machines or small VMs.
  • Some argue performance “shouldn’t matter” for a harness that mostly waits on LLM calls; others reply that responsiveness, battery life, and the ability to run many agents or on tiny instances are practically important.

Architecture & Features

  • Design choices for small footprint: Rust instead of JS/Python, load-on-demand connectors, smallvec/compactstring, LTO, size-optimized builds, single-threaded async.
  • Integrated features include prompt library (markdown-based “agents”), git worktrees, and “Ralph Wiggum” looping for long tasks.
  • No Skills or Subagents yet; single context buffer to stay lean, but subagents are under consideration.

Skills vs Prompt Library

  • One side says Skills are more than prompts: they enable runtime discoverability via metadata, multiple skills per request, and context isolation without resetting system prompts.
  • Others argue a prompt library is a simpler alternative for many use cases, trading automation and caching efficiency for transparency and minimalism.
  • Concerns are raised that changing prompts mid-session breaks caching and increases token cost, and that the current Anthropic implementation may not use cache-control.

Extensibility & Scripting

  • Some see lack of a Pi-style extension system as a key limitation; they expect agents to self-extend and hook deeply into the harness.
  • Proposed approaches: wasm-based plugins, JSON-RPC tool processes, Lua or Rhai scripting; skepticism toward embedding heavyweight JS/Deno runtimes.

Rust vs Other Languages

  • Broad agreement that language alone doesn’t fix bad architecture, but many expect idiomatic Rust to beat TS/JS harnesses in memory and cold start.
  • Debate over whether Rust is really “low-level,” and whether the Rust enthusiasm is justified or just hype.

Security, Sandboxing & Reliability

  • Code was informally checked (including with LLMs) for hidden telemetry; no extra network calls beyond LLM/MCP are reported.
  • Panic was initially set to abort to save size, then reverted after feedback about lost stack traces.
  • Supports bubblewrap-based sandboxing; there’s consensus that sandboxing is important, with some arguing it should be the default and more configurable.