vLLM Agentic API 0.8 makes stateful open-model agents easier to operate

vLLM’s Agentic API project has released version 0.8.0, expanding the operational layer that sits in front of model servers for stateful agent workloads. The release adds opt-in OpenTelemetry support, keeps image inputs intact across response continuations and introduces Brave Search as another server-owned search provider.

What changed

The release can emit OpenTelemetry traces and HTTP metrics over OTLP HTTP/protobuf, propagate W3C trace context and correlate those traces with local logs. Telemetry remains disabled by default, so operators must opt in rather than absorb a new observability path during an upgrade.

Agentic API 0.8 also preserves images across Responses API requests, stored continuations and both HTTP and WebSocket transports. The project says it now resolves image capabilities consistently for Codex model catalogs and launchers. That closes an important continuity gap for agents that analyze images across multiple turns instead of treating every image request as a disconnected inference call.

The tool layer gained Brave Search as a selectable web-search provider, including provider concurrency limits. MCP discovery now follows pagination cursors returned by tools/list, avoiding an incomplete tool inventory when a server exposes more than one page. The release also separates logical retained-response accounting from raw SSE framing and reports token usage across every round of a Messages tool loop.

Who should care

Platform teams using vLLM behind coding assistants or other agent clients get the clearest benefit. The gateway’s role is not faster token generation; it is to own conversation state, streaming, tool execution and continuation around an inference backend. Version 0.8 makes that control plane more observable and less likely to lose multimodal context during longer sessions.

Developers embedding Agentic API as a Rust dependency should review the migration notes. The release changes source-level APIs around input content, configuration, typed resource-limit errors and web-search provider construction. Existing persistent Codex session homes should also be regenerated so that the updated image capability data is picked up.

What to do

The project publishes 0.8.0 through crates.io and PyPI, with portable Python wheels for Linux x86_64 and both Intel and Apple Silicon macOS. A verified Docker image is available as vllm/agentic-api:v0.8.0.

Operators should enable telemetry first in a test environment, verify their collector and trace-context settings, and exercise stored multimodal continuations before promoting the release. Teams maintaining Rust integrations should treat this as a migration rather than a drop-in library update.