Databricks Introduces Omnigent Beta for Agents That Run Across Coding Tools

Databricks has introduced Omnigent, a managed beta that lets engineers define an agent once and run it through Claude Code, Codex, or direct APIs. Nimble is positioned as the web-search provider, turning external retrieval into a choice made in the agent setup rather than one inherited from each runtime.

One agent spec, several ways to run it

The product is a layer above the agent harness, the environment that supplies the model and tools. Databricks says an Omnigent definition can retain its model, tools, policies, and limits while the underlying harness changes. Its documentation also lists Cursor, Pi, and custom agents among the supported paths.

For agents running on Databricks-hosted models, calls go through the company’s Foundation Model APIs. Databricks says that places cost tracking, audit records, and governance in one place. The managed version requires Omnigent beta enablement for a workspace and a region that supports Unity AI Gateway.

Web search becomes part of the spec

Nimble can be selected through Omnigent’s built-in web-search tool. Nimble says its integration offers a light mode for broad discovery and a deep mode that extracts page content; it also supports settings for the number of results returned. Databricks describes Nimble Web Search Agents as handling multi-source searching, extraction, cross-checking, and citation generation for research and enrichment work.

tools:
  builtins:
    - name: web_search
      search_provider: nimble

A performance case that still needs validation

Databricks cites Nimble’s testing, which reported that adding Nimble web search increased LLM benchmark accuracy from 46% to 71% and cut web-search costs in half. Those are vendor-reported results, not an independent comparison, and Databricks does not provide a methodology or task breakdown in its post.

The immediate appeal is operational: teams can keep the same agent definition while changing its harness or model, and can set one search provider across those runs. But the claimed quality and cost advantage remains a result for prospective users to test against their own workloads.