Tetriz Harness

Product Introduction

1. Definition: Tetriz Harness is an open-source, self-hosted platform for managing and optimizing the configuration (the "harness") of AI-powered coding assistants like Claude Code, Cursor, and Kiro. Technically, it is a suite of Go-based microservices and libraries that function as an AI Agent Operations (AIOps) and Configuration Management system.
2. Core Value Proposition: It exists to solve the operational chaos of managing AI coding agents at scale. It provides systematic harness recommendations to identify broken or missing rules, skills, and permissions, and a control plane to safely distribute and roll out approved harness changes to specific developers or teams, enabling data-driven improvement of AI-assisted development workflows.

Main Features

1. Harness Recommendations Engine: This feature analyzes an organization's codebase and coding session data to suggest improvements to AI agent configurations. It works through two primary scan modes: a static DETECTED scan that parses committed harness files (like .cursorrules, claude_desktop_config.json) to find conflicts, missing file references, and over-broad permissions. The more advanced DERIVED engine uses session analyses, code-review themes, and security signals (processed via the bundled tetriz-py service with an LLM) to propose new rules, agents, and hooks tailored to the team's actual coding patterns.
2. Harness Control Plane: A centralized distribution system for AI agent artifacts. It consists of a Go server library, a reference admin UI, and a harness-agent CLI. It allows administrators to target specific artifacts (skills, rules) to named individuals, teams, or the entire organization. The system installs files into the correct user configuration directories or repository checkouts and provides real-time reporting on installation status across the fleet.
3. Test/Control Rollout Capability: A critical operational feature built atop the control plane. It enables manual A/B testing of harness changes by allowing admins to target a "test" group with a new artifact while leaving a "control" group unchanged. Outcomes are measured externally (e.g., via session analysis), and the change can be rolled out broadly or withdrawn based on the results, supporting evidence-based iteration of AI agent behavior.

Problems Solved

1. Pain Point: Unmanaged, fragmented AI agent configurations lead to inconsistent code quality, security risks from over-permissive tools, and wasted potential as agents operate without team-specific context. Manually distributing and versioning harness updates is error-prone and slow.
2. Target Audience: Engineering Managers and Developer Productivity Teams at companies scaling AI-assisted development; Platform Engineers responsible for internal developer tools and security; Individual Developers or Small Teams seeking to systematically improve their personal or shared AI coding workflows.
3. Use Cases: A team adopting Cursor wants to enforce consistent code style and security rules across all members. A platform team needs to deploy a new, company-specific coding skill to the frontend team only for testing. An engineering manager wants to identify why AI agents are generating errors in a specific part of the monorepo and push a fix.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros:
     • Open-Source & Self-Hosted: Provides full control over data (session data, codebase context) and avoids vendor lock-in.
     • Context-Aware Recommendations: Moves beyond generic rules by deriving suggestions from the actual repository and team coding signals.
     • Precise, Auditable Deployment: Granular targeting and installation reporting provide clear audit trails for compliance and debugging.
     • Designed for Experimentation: Built-in support for test/control rollouts is unique and essential for safe, iterative improvement.
   • Cons:
     • High Implementation Complexity: Requires significant setup, including running services (tetriz-py with an LLM key), building ingestion adapters for session data, and self-hosting infrastructure.
     • Dependent on External Signals: The most powerful DERIVED recommendations require populated "session signal" tables, which need integration with tools like Agent Desk or custom pipelines.
     • Not a Turnkey SaaS: Lacks a managed cloud service; the operational burden is on the user.

2. Key Alternatives & Differentiation:

   • Manual .cursorrules Management: Teams share rules via docs or copy-paste.
     • Differentiation: Tetriz automates discovery of issues, proposes data-driven improvements, and provides a reliable distribution mechanism, eliminating sync errors and enabling measured rollouts.
   • Generic Configuration Management (Ansible, Chef): These tools can push files but lack AI/agent-specific context.
     • Differentiation: Tetriz understands the semantics of a "harness" (rules, skills, hooks), can analyze their efficacy against code, and integrates with the AI coding agent ecosystem natively.
   • Commercial AI Coding Platform Rules Engines: Some platforms offer basic rule management within their walled garden.
     • Differentiation: Tetriz is agent-agnostic (works across Claude Code, Cursor, etc.), offers deeper codebase-integrated analysis, and gives organizations full ownership of their optimization logic and data.

Frequently Asked Questions (FAQ)

1. What is an AI coding harness? An AI coding harness is the complete set of instructions, rules, custom skills (tools), system prompts, and configuration files that guide and constrain the behavior of an AI-powered coding assistant like Claude Code or Cursor, shaping how it interacts with your codebase.
2. How does Tetriz Harness get data about my team's coding sessions? Tetriz does not collect data directly. It relies on you to provide data through an ingestion adapter. It is designed to pair with tools like Agent Desk, which can inventory local harness files and extract anonymized session events, which you then process and feed into Tetriz's recommendation engine.
3. Can I use Tetriz Harness with just one AI coding agent, like Cursor? Yes. While it supports multiple agents (Claude Code, Codex, Cursor, Kiro), you can deploy it for a single agent. Its static analysis (DETECTED scan) and control plane distribution work effectively for managing rules and skills for one agent type.
4. Is Tetriz Harness a cloud service or self-hosted software? Tetriz Harness is primarily self-hosted, open-source software. You run its Go services and UI on your own infrastructure. This ensures all sensitive codebase and session data remains within your environment.
5. What is the difference between Harness Recommendations and the Control Plane? Harness Recommendations is the analytical half: it scans and analyzes to suggest what to change. The Control Plane is the operational half: it is the system that deploys and manages those approved changes across your developer fleet. They are designed to be used together but can also function independently.