Agentic AI Platform Change Log

Observability / auditability

Galileo, now sold as Splunk Agent Observability, released Luna-3, the third generation of its own evaluation models, for enterprise users in the us0, us1, eu0 and au0 realms. Luna-3 accepts images as inputs for supported evaluators and uses the same logical inputs as LLM as a judge evaluators across sessions, traces and spans. The vendor says it is 30 to 40% faster than the previous generation, with 8 to 10% better accuracy on Prompt Injection, Toxicity, Sexism, PII and Context Adherence.

Luna models are the cheaper way to score all production agent traffic, so a faster model with image support lowers the cost of continuous guardrail checks and covers agents that handle images. The speed and accuracy gains are vendor figures without a published method.

Funding / partnership

Galileo has officially transitioned to become Splunk Agent Observability. The existing platform and documentation now only apply to legacy customers who onboarded prior to this date, while new users are directed to Splunk's dedicated infrastructure.

Buyers evaluating Galileo must now engage with Splunk for new deployments and refer to Splunk's ecosystem. This signals a major shift in product ownership, support channels, and future roadmap alignment.

Human approval / guardrails

Galileo released Annotation Queues to general availability, alongside updates to OpenAI model integrations and enhancements to the AI Assistant beta. The release also introduces flexible custom model integrations for connecting proprietary or third-party models into observability workflows.

The general availability of Annotation Queues provides a stable, production-ready mechanism for human-in-the-loop feedback and ground-truth dataset building. Expanded model support allows teams to evaluate a wider variety of proprietary and third-party LLMs.