LangGuard launches Arbiter for agent runtime governance

LangGuard announced general availability of LangGuard Arbiter at the Databricks Data and AI Summit in San Francisco. The new engine is designed to enforce deterministic, real-time control over agent actions before they reach enterprise systems, as AI agents spread across more tools and workflows.

Why it matters: - Enterprises are adopting agents faster than they can define and enforce oversight for the actions those agents take. - LangGuard Arbiter is meant to close that runtime authority gap by deciding whether an agent action is allowed, blocked, or escalated before the action reaches production systems. - The launch follows a high-profile April 2026 incident in which an agent deleted a customer database in nine seconds after reasoning past a credential mismatch to a destructive API call.

What happened: - LangGuard announced general availability of LangGuard Arbiter on June 10, 2026, at the Databricks Data and AI Summit. - The company positions Arbiter as a deterministic enforcement engine for agent actions. - LangGuard also said current Databricks customers can start a 14-day free trial of the full platform beginning June 15. - The company plans to show the product at booth #727 at the Databricks Data and AI Summit at Moscone Center from June 15-18.

The details: - LangGuard GRAIL Data Fabric, launched earlier this year, is designed to provide visibility into every action agents take. - LangGuard Arbiter operates at the agent action surface, the layer between an agent’s reasoning and the enterprise systems it can reach. - Every attempted action is evaluated before it reaches the target system. - Arbiter returns one of three outcomes: ALLOW, BLOCK, or ESCALATE to a human authority. - Every policy generated by Arbiter is red-teamed against adversarial agent behavior before acceptance. - Policies that pass are added to a verified policy ledger with a provenance chain from compliance intent to enforcement. - LangGuard says that ledger is the compliance artifact auditors need. - The LangGuard Platform includes three components: GRAIL Data Fabric, SCOPE-MCP server, and Arbiter. - SCOPE-MCP maps what agents can reach at design time. - LangGuard says the platform is built on Databricks and supports any agent harness. - The company says the platform can be deployed as an app in a Databricks workspace and connected to an existing agent harness within minutes. - LangGuard is headquartered in Austin, Texas, with offices in the Bay Area and Canada. - LangGuard says it is a Databricks-native app built for AI builders, Forward Deployed Engineers, IT, compliance teams, and chief data and AI officers. - LangGuard’s website is more information. - The company’s social channels listed in the release include LinkedIn, YouTube, and X.

Between the lines: - The release frames agent governance as a new enterprise control category, not just a monitoring problem. - LangGuard is betting that visibility alone is not enough, and that enterprises will want enforcement at the moment of action. - The focus on SOX, GDPR, segregation of duties, and the “lethal trifecta” suggests the product is aimed as much at compliance and security teams as at AI builders.

What's next: - LangGuard is offering a 14-day trial to current Databricks customers starting June 15. - The company is using the summit to recruit enterprise buyers and introduce the platform to teams already building with agent harnesses. - The broader test will be whether enterprises adopt real-time policy enforcement as agent use moves from experimentation into production.