Harness
The runtime that executes an agent loop: prompts, model calls, tools, files, commands, and streamed output.
What Omnigent is, why a meta-harness exists, and how it differs from a single agent harness or SDK.
documented
Omnigent is an open-source meta-harness: a common layer above agent harnesses such as Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, SDK agents, and custom YAML agents. Its job is not to replace every agent. Its job is to make sessions, policy, sandboxing, collaboration, interfaces, and deployment portable across those agents. Sources: Omnigent README, coding agents docs, Databricks overview.
The runtime that executes an agent loop: prompts, model calls, tools, files, commands, and streamed output.
A layer above harnesses. It wraps multiple agent runtimes with one operational surface for sessions, governance, and collaboration.
When docs do not clearly describe a feature, this guide marks it as inferred or under-documented rather than treating it as fact.
The Databricks launch post frames the problem as agent work becoming multi-agent and multi-tool: teams have several agents open, copy context between them, and rebuild governance around each new runtime. Omnigent moves the shared operating layer above the individual agent. Source: Introducing Omnigent.
An agent SDK helps you build one agent or one agent app. Omnigent can run SDK-based and YAML-defined agents, but the platform idea is broader: persistent sessions, terminal/web/mobile/desktop interfaces, team sharing, REST APIs, contextual policies, and OS/cloud sandboxing. Sources: custom agents, web UI, policy overview.
The harness docs distinguish Direct mode, where Omnigent drives the model and tools itself, from Native TUI mode, where Omnigent boots a vendor’s terminal UI and mirrors it back while adding collaboration and policy wrapping. Supported harnesses documented in the table include Claude Code, Codex, Cursor, Antigravity, Goose, Qwen Code, Kimi, Hermes, Pi, OpenCode, Kiro, Copilot, and OpenAI Agents SDK. Source: harnesses.