agentsmith
worthwhileA structured, battle-tested agent operating model that enforces verification, handoff, and system evolution — worthwhile for teams, overkill for one-off
What it is
Agentsmith is a comprehensive, installable set of rules, profiles, and scripts that governs how an AI coding agent (Claude Code, Codex, Gemini) operates in a project. It enforces a unit-of-work session model with planning, verification, and handoff, plus a system-evolution mindset for continuous improvement of the rules themselves.
How it differs from vanilla Claude
A vanilla Claude Code can follow a CLAUDE.md but lacks built-in verification gates, a structured handoff protocol, an anti-rationalization table, and a process for evolving its own instructions. Agentsmith provides these as a reusable, tested harness that has been iterated over real projects. It also includes scripts (verify.sh, handoff.sh) that the agent can invoke, which a base agent wouldn't have by default.
Skill, plugin, or workflow shift?
workflow-shift because adopting Agentsmith changes the entire operating model: sessions become atomic work units, 'done' requires real verification, handoffs preserve context via recall prompts. It's not a simple plugin or skill; it's a new set of conventions and practices that both the agent and the human follow. The integration is deep and affects how the project is structured.
Devil's advocate — is this just complexity?
A disciplined developer can already achieve similar results by writing a good CLAUDE.md with the same principles (verify before done, handoff procedure, etc.). Agentsmith packages these ideas with extra scripts and documentation, but the core concepts are not new — they're standard engineering discipline. The repo has many files and may be over-engineered for what could be a 50-line prompt. Moreover, as base agents like Claude Code gain built-in verification and memory, the harness may become obsolete. The '90% harness, 10% model' claim is a marketing framing, not a measured fact. For a solo developer on simple projects, this is unnecessary complexity.
What would make it better
A lighter, single-file version with just the core rules for quick adoption. Tighter integration with agent SDKs to avoid relying on shell scripts. Automated profile detection based on project files (e.g., package.json, Dockerfile). Built-in metrics to measure how often verification gates catch issues. A visual dashboard for monitoring agent sessions across projects.
The honest case for it
For a team doing serious AI-assisted development across multiple projects, Agentsmith saves weeks of trial and error. It codifies hard-won lessons: verification gates that actually run your tests, handoff that recovers context exactly, a process for improving rules as failures happen. The documentation is thorough enough to onboard new engineers. It's not just a prompt — it's an engineering discipline, transferable across projects and models. The layered setup (global core + per-project profile) is elegant and scales well.
Who it's for
Audience fit
Depth and leverage for a technical engineer who wants to understand it and level up their workflow — not just offload work.
Value for someone who wants a more capable tool without the technical depth — accessible, does-it-for-you.
Designed for engineers who want disciplined, repeatable agent workflows. Vibe coders may find the rules restrictive and the setup overhead not worth it for quick demos.