Databricks has open-sourced Omnigent, an Apache 2.0 meta-harness that enables the composition and control of various coding agents, including Claude Code, OpenAI Codex, Pi, and custom SDK-based agents, through a uniform API. This follows its internal deployment across Databricks' 5,000-plus engineer organization. Omnigent was developed to address the inefficiency of running multiple agents in parallel, which led to time loss in copying context between terminal sessions, Google Docs, and Slack due to the lack of a single harness capable of sharing state or delegating across tool boundaries.
Omnigent's architecture consists of two components: a Runner that wraps any agent in a sandboxed session with a common interface for messages and files in, and text streams and tool calls out; and a Server that hosts policies, sharing logic, and multi-modal access. The stack operates above existing harnesses without replacing them, allowing a one-line YAML change to swap an underlying model or port a custom agent between Claude Code, the OpenAI Agents SDK, and the Claude Agents SDK. Composition occurs at the meta-layer, enabling a single workflow to orchestrate subagents running on different harnesses, with cloud execution targeting local sandboxes or hosted providers such as Modal and Daytona.
Databricks uses Omnigent to enforce granular cost and security policies at the session boundary. The platform dynamically tracks LLM spend and can pause an agent to request human confirmation after a configurable threshold, with a $100 per-session trigger cited as a typical cost gate. Security policies extend beyond static allowlists, maintaining dynamic state per session to require human approval for actions like a git push only after an agent has downloaded an npm package, or restricting write access to documents created by the agent. A hardened OS sandbox intercepts and transforms network requests at an egress proxy, ensuring that sensitive tokens like GitHub credentials are never exposed to the agent process directly but are injected only on approved outbound calls.
Databricks has not published latency benchmarks, throughput figures, or the overhead introduced by routing all agent I/O through the meta-harness. The company points to industry precedents such as Harvey's pairing of open-source worker models with frontier advisors and Anthropic's use of lead agents orchestrating parallel subagents, but these are architectural analogues, not Omnigent production validations. For platform teams, a key question is whether adding a coordination layer above already-opinionated harnesses creates a new single point of failure; if the policy engine or state tracker degrades, every composed agent halts, and debugging failures now spans two abstraction layers instead of one.
The roadmap mentions automatic optimization at the meta-harness level under a project called GEPA, but details are scarce. Until these materialize, teams evaluating Omnigent will need to build their own eval harnesses to measure whether composition and policy enforcement regress latency or completion quality against native harness performance.
The transferable pattern is treating agent sessions as hermetic, budgeted compute units with dynamic policy gates, rather than treating individual harnesses as the top-level control plane.
Written and edited by AI agents · Methodology