Six months into the era of AI coding agents, a strange problem is emerging: every agent behaves differently. Claude Code has one personality. Codex has another. Gemini brings its own assumptions. Teams that switch between models find themselves rewriting instructions, retraining behaviors, and battling inconsistent output. AgentSmith, a new open-source project from PromptPartner, proposes a different approach: instead of adapting to each agent's native behavior, force all agents to run on the same operating system. Launched just two days ago, the project has already accumulated 307 GitHub stars and is growing fast as developers discover that the harness around the model matters more than the model itself.
What AgentSmith Does
AgentSmith is a portable, model-agnostic agent operating harness that standardizes how AI agents behave across different models. It works by providing a universal core of rules, tools, memory structures, guardrails, and feedback loops, combined with swappable work-type profiles for specific tasks. The core insight is what the project calls the 90/10 rule: the model accounts for roughly 10 percent of the outcome, while the harness (the rules, tools, memory, guardrails, and feedback loops around it) accounts for the other 90 percent. AgentSmith aims to make that 90 percent reusable across any agent that reads a CLAUDE.md, AGENTS.md, or GEMINI.md file. It supports profiles for software development, devops and setup, marketing and outreach, document creation, data crunching, research, design, and general admin. Each profile adapts the core harness to a specific type of work without changing the underlying discipline. The harness itself was not designed in a vacuum. It grew over roughly six months of real autonomous work across data crunching, marketing outreach, and software development projects, going through five iterations before being scrubbed of project-specific details and released as a universal core with swappable profiles. This real-world provenance matters because it means the harness has been tested against actual production use cases, not theoretical workflows.
How to Use It
Getting started takes one command. Clone the repo and run the setup script, or if you are already using a coding agent, paste a single prompt into your agent's chat and it installs itself autonomously. The setup wizard auto-detects the right profile for your project, asks a couple of quick questions, and assembles the harness. Each session follows a plan-implement-verify-ship cycle with real autonomy. The agent decides routing and scope, and only pauses for genuinely human-needed inputs like a missing credential or an external surprise. Nothing ships on vibes: the verify step runs real project checks, bug fixes require a failing test first as proof, and review gates catch what tests cannot. At roughly 25 to 30 percent context usage, the agent automatically safe-states and writes a recall prompt so the next session resumes cleanly without losing progress. The project ships with 19 files including comprehensive documentation, a setup script for Unix systems, a PowerShell setup script for Windows, and detailed guides covering everything from your first hour with the harness to troubleshooting common issues. The documentation is organized as a numbered series of markdown files that serve as both reference and reading guide.
Comparison to Alternatives
AgentSmith enters a field that has several existing approaches but no dominant standard. Many teams write custom CLAUDE.md or AGENTS.md files for each project, but these are typically one-off efforts that do not transfer between projects or models. Some organizations use curated prompt libraries or template repositories, but these lack the structured verification gates and session management that AgentSmith provides. The closest comparable tool is probably the built-in agent configuration systems from each model provider: Anthropic's CLAUDE.md system for Claude Code, Google's AGENTS.md for Gemini, and OpenAI's custom instructions for Codex. AgentSmith's key differentiator is that it wraps all of these in a single, consistent interface. You install the harness once and it works the same way regardless of which model you are using underneath. For teams that switch between models based on pricing, capability, or availability, this consistency is the primary value proposition. No other open-source project currently attempts to provide a unified agent operating system across model providers at this level of completeness.
Why It Matters for Founders and Developers
AgentSmith addresses a pain point that has been silently growing as AI coding tools proliferate. The conversation for the first half of 2026 has been dominated by individual agents: Claude Code versus Codex versus Gemini. But teams working across these tools have discovered that each agent has its own quirks, its own prompt format, its own way of handling context, and its own failure modes. Switching agents means retraining not just the model but the entire workflow. AgentSmith's model-agnostic approach is the first serious attempt to standardize agent behavior across providers. This is the same pattern that played out in the operating system wars: individual applications matter less than the platform they run on. If AgentSmith succeeds, it could become the Linux of AI agent infrastructure: an open standard that lets developers swap models without swapping workflows. For solo founders and small teams who cannot afford to rewrite instructions every time they try a new model, that portability is a significant productivity gain. The project is also notable for what it reveals about the maturity of the AI coding tool ecosystem. Six months ago, the conversation was about whether AI coding agents worked at all. Now it is about how to standardize their behavior across providers. That shift signals that the technology has passed an important threshold.
Who this is for: Developers using AI coding agents who want consistent behavior across models. Teams evaluating which agent to standardize on. Anyone tired of rewriting CLAUDE.md files every time they try a new coding assistant. Solo founders who switch between Claude Code and Codex depending on the task.
The project is MIT-licensed and available now on GitHub at github.com/PromptPartner/agentsmith. The setup script accepts both interactive wizard mode and headless flags for automated deployment, and the full documentation is included in the repository's docs directory.