Databricks launches open-source Omnigent for AI agents

Databricks has launched Omnigent, an open-source meta-harness for AI agents designed to let organisations manage different agent systems through a single framework.

Omnigent sits above existing agents and provides an orchestration layer for teams using multiple models, frameworks and development environments. It has been released in alpha under the Apache 2.0 licence.

Businesses are beginning to use AI agents for workflow automation, software coding, research and decision support. That shift has created a fragmented landscape, with teams often running separate agent systems that do not work together easily.

Omnigent is intended to address that problem by linking different agents through a common interface. Developers can integrate agents built with tools including Claude Code, Codex, Pi and in-house frameworks.

Three functions

The product groups its features into composition, control and collaboration. Composition allows teams to combine agents, models and frameworks, and to move between agent systems with limited code changes.

Control focuses on governance, security and cost management. The system can apply stateful policies that track agent actions and enforce guardrails beyond prompt-based controls.

Collaboration is aimed at teams working together on agent-led tasks. Users can share live agent sessions, review outputs together and guide workflows in real time.

Additional features include cloud execution environments, operating system sandboxing, contextual security policies and cost budgeting controls. The framework also supports custom multi-agent systems built across different software environments.

Internal use

Databricks linked the launch to its experience using AI agents inside its engineering organisation. It said Omnigent drew on lessons from internal deployments and from building agent-based products including Genie.

That work led Databricks to conclude that more advanced AI workflows increasingly involve several models, several agent frameworks and several users at the same time. In that setting, it argues, a higher-level coordination layer is needed to govern interactions across tools and teams.

Releasing Omnigent as open-source software may also help Databricks widen its reach among developers already using different agent tools. By supporting external frameworks as well as custom-built systems, the company is positioning the software as a neutral management layer rather than a tool tied to a single model or vendor stack.

Wider market

The launch comes as technology companies race to define the software layer around AI agents, which are becoming a focus of enterprise IT spending. While much attention has centred on foundation models, a growing part of the market is shifting towards tools that manage how those models are deployed, supervised and connected to business processes.

Governance has become a particular concern as organisations move from experiments to live deployments. Companies adopting agents are under pressure to set rules around security, user access, budget limits and auditability, especially when several systems interact with one another.

Databricks said Omnigent was built to provide those controls in context rather than relying only on prompts or fixed instructions. That approach suggests a push towards software that monitors agent behaviour over time and applies rules based on actions and environment as well as user input.

The release also reflects a broader move towards multi-agent systems, in which different specialised agents are assigned separate tasks within a wider workflow. Supporters of that model say it can improve flexibility and allow companies to combine coding assistants, research tools and domain-specific agents without rebuilding everything around one platform.

Databricks serves more than 20,000 organisations worldwide and says 70% of Fortune 500 companies use its products. Omnigent is now available in alpha as an open-source project.