Splunk defines eight levels of AI agent autonomy to mitigate failures

The shift from deterministic workflows to autonomous agents introduces significant operational risks that traditional software testing cannot address. By categorizing autonomy levels, streaming infrastructure providers can better secure attack surfaces like the Vertex AI Memory Bank or OpenAI Agents SDK against prompt injection and data poisoning. As multi-agent systems become more common, the 'coordination tax' and hallucination snowballs—where errors survive undetected through multiple stages—will require full-trace observability rather than simple output logging. This framework forces a transition toward rigorous, real-time evaluation of reasoning steps. Watch for the adoption of the A2A protocol by the 150 supporting organizations to see if standardized cryptographic identities can reduce environment-level security failures.

Splunk has been building its agent observability capabilities as part of a broader push to make AI agent monitoring a first-class product category. In March 2025, Splunk launched Agent Observability as a dedicated feature within its platform, targeting multi-step AI agent workflows that require tracing across reasoning chains, tool calls, and memory stores. The product integrates with OpenTelemetry-based instrumentation and supports the Model Context Protocol for capturing agent-to-tool interactions. This positions Splunk alongside Datadog and New Relic, both of which have shipped LLM observability modules, but Splunk's taxonomy approach is distinctive in that it maps failure modes to specific autonomy levels rather than treating all agent errors as equivalent. The business and standards landscape around agent observability is consolidating quickly. The Linux Foundation announced in April 2025 that it would host the Model Context Protocol as an open governance project, giving MCP a neutral stewardship structure that enterprise buyers increasingly require before adopting protocol-dependent tooling. Google's A2A protocol, which Splunk references for agent-to-agent coordination, gained over 150 supporting organizations within weeks of its April 2025 launch, including Salesforce, SAP, and ServiceNow. For streaming infrastructure teams evaluating these protocols, the governance question matters: MCP under the Linux Foundation provides a clearer path to vendor-neutral implementation than proprietary agent frameworks. On the technical side, competing observability approaches for AI agents are emerging from both APM vendors and AI-native startups. Datadog released LLM Observability in general availability during 2025, offering trace-level visibility into prompt-response pairs and token costs, while LangSmith from LangChain provides evaluation-driven monitoring specifically for chain and agent architectures. For streaming platforms that deploy AI agents in content recommendation, ad decisioning, or automated QC pipelines, the choice between these tools depends on whether the team needs general-purpose APM correlation (favoring Datadog or Splunk) or deep evaluation-loop integration (favoring LangSmith or Anthropic's own evaluation models). Splunk's eight-level taxonomy gives engineering teams a shared vocabulary for deciding which observability depth is required at each autonomy tier, reducing the risk of AI agent deployment risks in production streaming workflows, especially as agentic video intelligence architecture gains traction, while AI agent video security remains a critical consideration for content pipelines, alongside new tools like the Open Agent Safety Platform.