Every minute an engineer spends copying output from Claude Code into Codex, then pasting context into Pi, is a minute spent doing the job the agent was supposed to do.

Five agents open, zero shared memory between them

Engineers running multiple AI coding agents today have no shared layer between them. Each harness holds its own session, its own context, and knows nothing about what the others are doing.

One orchestrator now sits above all of them

Databricks Open wraps terminal coding agents like Claude Code, Codex, and Pi, plus SDKs like OpenAI Agents and the Claude Agents SDK, behind a single uniform API. You start a session with one command, and it immediately surfaces in your terminal, a local web UI at localhost:6767, and a web API simultaneously. Multiple harnesses become interchangeable workers under one orchestrator, and the server layer enforces policies and controls sharing across all of them.

Platform engineers are the first ones this changes things for

Anyone managing multi-agent workflows across a team will feel this immediately:

• Platform engineers who maintain separate agent environments per project and spend hours reconciling outputs
• AI engineers who need a governance layer before deploying agents to production but currently have none
• Engineering leads who want to swap one coding agent for another without rewriting the surrounding infrastructure

The composability angle matters as much as the coordination. Because each harness is treated as an interchangeable part, switching from Codex to Claude Code mid-project does not require rebuilding the session logic around it.

The agent tooling market just made a structural bet on orchestration

Microsoft’s Copilot and Google’s Gemini both push single-vendor agent stacks, which means any team running mixed tooling has been left to build their own glue. Databricks shipping this under Apache 2.0 with Neon signals that cross-agent orchestration is moving from internal tooling into open infrastructure, and teams that build on proprietary wrappers now have a real alternative.

What you can actually do with it today

• Run Claude Code, Codex, and Pi from one shared session interface
• Swap coding agents without changing the orchestration layer above them
• Expose any active agent session through a web API for team access
• Apply governance policies across all agents from a single server config

Open source under Apache 2.0, built with Neon, free to self-host.

Omnigent requires you to supply your own models and infrastructure. Teams without existing cloud infrastructure will need to build that foundation before the orchestration layer adds much value.

LangGraph covers multi-agent orchestration for teams already inside the LangChain ecosystem. CrewAI takes an agent-role framing rather than a harness-wrapping approach, which suits different architectures depending on whether your team thinks in workflows or workers.

The race to own the layer above AI agents just went open source

The most important AI infrastructure decision in 2025 may not be which model you pick, but which orchestration layer sits above all of them. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.