Show HN: Axe – A 12MB binary that replaces your AI framework

Original Article ↗ Hacker News Discussion ↗

Overall Concept and Positioning

  • Axe is a small Go binary that treats LLM “agents” like Unix commands: each agent is defined in TOML, does one focused job, and is composed via the shell (pipes, files, scripts).
  • It is not an LLM itself but a runtime that calls LLM APIs (including OpenAI-compatible ones, and others like Gemini planned).
  • Compared to other tools:
    • Different from in-code prompt templates (e.g., dotprompt) because it’s a CLI execution layer, not an app framework.
    • Closer to shell-based “tiny agents” or claude -p workflows, but with a more structured agent/config and chaining model.
    • Seen as more automation-oriented than interactive tools like sgpt, Pi, Codex, or Claude’s full IDE integrations.

Use Cases and Pipelines

  • Several multi-step flows described:
    • YouTube transcript → blog-style post → Instapaper.
    • Voice notes → related note lookup → blog draft → cleanup (e.g., remove em dashes).
    • “Ghost blog” with each step saving artifacts to files between agents.
  • People propose:
    • Code refactoring, test generation, build harnesses, and PR prep as a “junior dev in the repo.”
    • Git commit message generation.
    • RPG character creation workflows and invoice-like processing.

State, Artifacts, and “Memory”

  • Agents commonly pass data via files between steps; some users archive and summarize artifacts with git integration.
  • Session support is requested but not yet implemented; current use skews toward explicit file-based state.
  • Some enthusiasm for “persistent memory”; others criticize the term as vague and potentially overcomplicated, asking for clearer, minimal explanations.

Cost Control

  • Single-purpose agents are seen as helpful for keeping prompts small and behavior predictable.
  • Concerns raised about accidentally fanning out many agents and cost explosions; token or budget limits per agent are discussed but not yet fully designed.

Security and Safety

  • Prompt injection and destructive actions (e.g., with calendars, email, GitLab) are core worries.
  • Suggested mitigations:
    • Running agents in containers (e.g., Docker) and path-sandboxed file skills.
    • Network egress allowlists for outbound connections.
  • Some skepticism about relying on chroot or similar as true security boundaries; security primitives at this layer are still evolving.

Implementation Details and Trade-offs

  • 12MB binary size triggers debate:
    • Some see it as large compared to ultra-small Zig binaries.
    • Others note Go’s runtime overhead and consider 12MB acceptable or minor.
  • Config details:
    • Uses TOML; a few would prefer JSON5.
    • Default agent location under ~/.config feels wrong to some; they want per-repo agents for better versioning.
  • Mixed preferences on UX:
    • Some prefer chat interfaces for iterative planning, then Axe-like tools for deterministic execution.
    • Desire expressed for a chat-based helper that generates Axe configs and for a shared “Awesome Axe” agent catalog.

Conceptual Debates

  • Debate over very fine-grained agents for coding:
    • Critics argue human analogues would be inefficient.
    • Supporters counter that narrow LLM responsibilities increase reliability and trust.
  • Terminology disputes appear (e.g., “clankers” for AIs), with some finding it funny and others off-putting.