Summary

AgentSmith provides a universal, model-agnostic operating harness for AI agents, addressing the critical need for robust systems beyond the AI model itself. It offers a portable, battle-tested framework for disciplined, autonomous AI work across diverse tasks, enabling reliable development and maintenance of AI-driven products and workflows.

How It Works

The harness combines a lean, universal core of unchanging rules with swappable work-type profiles. Agents operate autonomously through a plan-implement-verify-ship cycle, emphasizing rigorous verification (verify.sh, failing tests first) and continuous system improvement by fixing harness flaws. This ensures reliable, evidence-based outcomes and facilitates clean session handoffs.

Quick Start & Requirements

Installation is flexible: via an existing coding agent, running the ./setup.sh wizard (or pwsh ./setup.ps1 on Windows), or direct flag usage. Prerequisites include a compatible AI agent (e.g., Claude Code) and bash with optional jq. Detailed setup and documentation are linked via AGENT-INSTALL.md and the docs/ directory.

Highlighted Details

• Model-Agnostic: Works with various agents via standard description files (CLAUDE.md, AGENTS.md).
• Layered Setup: Install the core globally, then add thin profiles per project for efficient management.
• Autonomous Workflow: Agents execute plan-implement-verify-ship cycles with built-in autonomy and verification gates.
• Bundled Skills: Includes utility skills like /handoff, /verify, and /harness-help.
• Configurable Safety: Offers cautious and trusted safety modes for agent permissions.
• Self-Updating: Facilitates keeping the harness current via ./setup.sh --self-update.

Maintenance & Community

Developed by Lukas Hertig (PromptPartner), the project emphasizes practical application. Direct community channels are not listed, but contact is available via promptpartner.ai or booking a consultation.

Licensing & Compatibility

Released under the permissive MIT license, allowing commercial reuse, adaptation, and distribution with attribution. It contains no vendored third-party code, avoiding upstream obligations, and is designed to be free of project-specific data or credentials.

Limitations & Caveats

Requires a compatible AI agent reading harness description files. The trusted safety mode presents risks for sensitive environments due to elevated agent permissions. Windows users need PowerShell 7+ and may need to adjust execution policies.