Databricks open-sources Omnigent meta-harness for cost-controlled multi-agent orchestration; addresses $500M/month AI budget overruns

Databricks open-sourced Omnigent, a meta-harness for orchestrating multiple AI agents under a single control plane, released June 13 under Apache 2.0 license and announced in early alpha. Omnigent wraps existing agent harnesses—Claude Code, OpenAI Codex, Inflection Pi, LangGraph, CrewAI—without modifying them, providing a unified API for composition, governance, and team collaboration. It operates at the orchestration layer above individual agent runtimes, enforcing stateful policies (cost budgets, filesystem access, network boundaries, human-in-the-loop approvals) outside the LLM prompt layer.

The motivation is acute: Uber spent its entire 2026 AI budget in four months as 5,000 unmonitored engineers ran Claude Code sessions, with heavy users reaching $500–2,000 per month. One unnamed enterprise spent $500 million in a single month before finance teams intervened. Gartner forecasts 40% of AI agent projects will be canceled by 2027 due to cost spirals and inadequate risk controls. Omnigent's cost control sits at the infrastructure layer: it pauses sessions when spending thresholds are exceeded and requires human confirmation before proceeding. Support for multi-agent coordination enables teams to compose specialized agents (planner, implementer, reviewer, validator) in workflows with shared context and approval gates.

Omnigent integrates with Unity AI Gateway, Databricks' runtime governance layer announced at the summit. Gateway enforces centralized cost controls (spend visibility across providers, hard spend caps, smart routing), access policies, and unified observability for all agent interactions. Integration partners include Palo Alto Networks (Prisma AIRS for real-time inspection) and Zscaler (AI Guardrails for outcome-based controls). The architecture separates agent harnesses from policy enforcement: rules live outside the prompt and cannot be reasoned around because the sandbox enforces constraints before any LLM receives input.

For practitioners: Omnigent is production-alpha; evaluate in test environments before critical deployment. The out-of-prompt governance model is architecturally sound—immune to prompt injection because constraints are enforced at the orchestration layer, not in the prompt. This resilience matters for teams rebuilding workflows after the Anthropic Fable 5 export ban: governance at the orchestration layer survives model-level disruptions. Monitor for: (1) stability releases and early adopter case studies, (2) support for additional harnesses (GitHub Copilot Workspace, Cursor, etc.), (3) managed Omnigent pricing vs. self-hosted. Cost visibility is now table-stakes for agent deployment at scale.