Interests

Introducing Omnigent: an Open Meta-Harness

Properties2
descriptionMatei Zaharia introduces Omnigent, Databricks' open-source meta-harness that layers composition, collaboration, and contextual control policies on top of existing agent harnesses.
tagsai, agents, llm, orchestration, open-source, databricks

4 min read

Databricks keynote where Matei Zaharia launches Omnigent, an open-source meta-harness layer that sits above individual agent harnesses (Claude Code, Codex, custom agents) and solves three cross-cutting problems: composition, collaboration, and control.

The Problem with Agent Harnesses

Zaharia frames the landscape: agents work, coding agents are everywhere, but the ecosystem is fragmented. Every agent has its own “harness” (the software around the LLM that connects it to files, UI, security). Claude Code is a harness. Codex is a harness. Any custom agent SDK is a harness. Everyone is building their own, and they don’t interoperate.

Three specific pain points:

  • Composing agents - power users run multiple coding agents simultaneously, but switching between different harnesses is painful. Patterns like loops (agents supervising each other), advisors (cheap model + expensive model directing), and parallel workers lack a common interface.
  • Collaborating across humans and agents - Zaharia shows his desktop: Claude Code, Codex, Slack, Google Docs all open, copy-pasting text between them. You can’t hand a session to another engineer with full context.
  • Control and security - prompt injection risk, costly runaway agents, permission prompts that are useless (“run this 100-line Python script? yes/no”). Current guardrails are too static.

Omnigent’s Architecture

A new layer above the harness level. Open source, anyone can plug in their own harnesses, workflows, tools.

Bottom layer: existing agents (Claude Code, Codex, + anti-gravity, coarser added since Saturday launch, OpenAI Agents SDK for custom agents).

Interface: common message-in/event-stream-out protocol with tool calling. Two components:

  1. Runner - wraps each agent, enforces security, sandboxes, monitors
  2. Optional Server (Docker + Postgres + OpenTelemetry) - central history, agent sharing, MCP/skills management, security policies, artifact sharing. Gives every agent a uniform web UI, native app, mobile interface, and REST API, regardless of the original harness.

No Databricks dependency - runs standalone at any scale.

Composition

Define multi-agent workflows with YAML. Demo shows: starting Claude through Omnigent, then forking the same session to Codex (same history, different harness), then launching “Polly” - a multi-agent supervisor that delegates subtasks to both Claude Code and Codex simultaneously. The supervisor receives results, compares them, and produces a combined output.

Collaboration

Live session sharing - invite another engineer into your agent session in real time. They see the same files, diffs, and comments, can send messages to the agent, and can take over the session. All history persists on the server.

Contextual Control Policies

The key innovation over existing agent security models. Current tools have static per-tool-call allow/deny/ask policies. Omnigent supports stateful contextual policies:

  • Budget limits - “don’t spend more than $5 on this task”, with granular sub-agent budgets
  • Context-aware permissions - if the agent accessed confidential documents, block external email; if only public docs, allow it
  • OS sandboxing to protect the host machine

The demo shows setting a 20 cent budget and the agent pausing for approval when it exceeds it, eventually hitting a hard stop at the configured limit.

Lessons worth keeping

  • The “meta-harness” concept (harness above harnesses) is Databricks’ bet on where agentic infrastructure goes next - an orchestration+governance layer that’s harness-agnostic
  • Contextual, stateful security policies (not just static allow/deny) are the next frontier for agent safety - Zaharia explicitly calls this out as missing from every current framework
  • Cost budgets as a first-class security concept (not just a billing concern) is a novel framing
  • Omnigent is fully open source, no Databricks dependency - deliberately trying to become the “Kubernetes of agent orchestration”
  • The demo launched Saturday and already had community additions within 48 hours (anti-gravity, coarser support)
  • What Are Agents? - Matt Pocock on the agent-vs-workflow distinction; Omnigent is effectively a meta-workflow engine
  • The Economics of AI Reasoning - reasoning cost tradeoffs directly relevant to the advisor-model patterns Zaharia describes

Sources

Graph View