OmniGent: A Meta-Harness for Unified AI Agent Orchestration

The Meta-Harness Architecture

OmniGent acts as a unified abstraction layer that sits above existing AI agent harnesses like Claude Code or Codex. By providing a consistent interface for messages, file access, and tool calls, it allows developers to treat disparate agents as interchangeable workers within a single session. The architecture consists of three primary components: a runner that wraps agents in a uniform sandbox, a server that manages shared history and state, and a deployment layer that supports local, Docker, or cloud-based execution. Because the session state lives in the meta-harness rather than the individual tool, users can switch interfaces—moving from terminal to web UI or mobile—without losing context or file history.

Multi-Agent Composition and Governance

OmniGent enables complex agent workflows through composition and enforced policy gates. Users can define agents via YAML files, specifying prompts, tools, and the underlying harness. The platform includes two built-in agents: 'Poly', which decomposes tasks and routes code to different vendors for cross-verification (e.g., Claude writes, Codex reviews), and 'Debbie', a brainstorming partner that facilitates debates between two models. Governance is handled via an approval proxy that intercepts tool calls, allowing developers to enforce cost caps, PII scanning, and file-access restrictions globally across all agents. This prevents 'YOLO' runs by ensuring that even if an agent is granted broad permissions, the meta-harness enforces safety policies before any action is executed.