This Meta-Harness Changes How You Run AI Agents

Key Concepts

• Meta-Harness: An orchestration layer that sits above individual AI agent harnesses to provide a unified interface, shared session, and centralized control.
• Harness: The infrastructure surrounding an AI model, including agent loops, tools, memory, and UI.
• Omni (by Databricks): An open-source (Apache 2.0) meta-harness designed to unify disparate AI agents under a single roof.
• Work Trees: Isolated environments where different agents perform tasks independently before merging results.
• Cross-Vendor Review: A methodology where code written by one agent (e.g., Claude) is reviewed by another (e.g., Codex) to eliminate internal biases.
• OS Sandbox: A secure, isolated environment that restricts agent access to specific files and network resources.

---

1. The Problem: Agent Fragmentation

Currently, AI users rely on multiple agents (Claude, Codex, Pi, etc.), each trapped in its own "box." These agents lack shared sessions, history, or visibility into one another. This forces users to manually copy-paste data between tools. Furthermore, building custom solutions on top of specific harnesses leads to "vendor lock-in," where upgrading to a better model requires a complete replumbing of the existing infrastructure.

2. The Solution: Omni (Meta-Harness)

Omni acts as a universal layer that sits beneath various harnesses. Because most AI harnesses share a common external interface (messages/files in, text/tool calls out), Omni treats them as interchangeable workers.

Technical Architecture:

• Agent Definition: Agents are defined via simple YAML files containing prompts, tools, and harness configurations.
• Runner: Wraps agents in a uniform sandbox session.
• Server: Manages search history, policies, Model Context Protocols (MCPs), skills, and artifacts. It uses Postgres for state management and is deployable via Docker, Railway, Fly, or Cloud Sandbox.
• Unified Interface: Provides a single session object accessible via terminal, web, mobile, or REST API.

3. Key Capabilities and Frameworks

Composition and Orchestration

• Polly (Tech Lead Agent): An orchestration agent that decomposes complex goals into subtasks. It delegates work to coding agents, manages work trees, and routes code to a different vendor for review.
• Debbie (Brainstorming Agent): A dual-agent setup (e.g., Claude + GPT) that provides side-by-side answers and features a /debate command where agents critique each other to reach a consensus.

Control and Security

• Gatekeeper Proxy: Every tool call passes through a gate (allow/deny/ask). Unlike simple prompt-based requests, this is enforced at the layer level.
• Secret Management: Agents never see raw secret keys. The layer injects them only during the approval proxy process, ensuring safety even in "YOLO" (unrestricted) modes.
• Policy Enforcement: Users can set cost caps, risk scores, and file/repo scopes, preventing unauthorized actions.

Collaboration

• Shared Sessions: Because the session lives in the meta-harness layer, multiple users can view or interact with the same agent session in real-time. This enables "code driving," where a teammate can watch, chat, or fork the conversation.

4. Practical Application: Workflow Example

1. Initialization: The user runs the Omni installation command and configures local API keys for various models (Claude, Codex, Ollama).
2. Task Submission: The user provides a prompt (e.g., "Create a web UI for image generation").
3. Autonomous Execution:
   • Implementation: The meta-harness assigns the task to a coding agent (e.g., Claude).
   • Verification: The meta-harness automatically routes the resulting code to a different agent (e.g., Codex) for an independent review.
   • Refinement: If the reviewer finds issues, the meta-harness sends the feedback back to the implementation agent for correction.
4. Monitoring: The user tracks the process via a unified web UI or terminal, with full visibility into token consumption and session costs.

5. Synthesis and Conclusion

Omni represents a shift from using isolated AI tools to managing a cohesive AI workforce. By decoupling the "harness" from the "model," it solves the issues of vendor lock-in and fragmented workflows. Its ability to enforce security policies, manage costs, and facilitate multi-agent collaboration makes it a significant development for both individual developers and organizations looking to scale their AI operations. As an open-source project, it provides a battle-tested framework for building complex, multi-agent systems that are both modular and secure.