n8n

Workflow builders now get guardrails against deprecated nodes, and the AI agent experience sees reliability fixes across the 2.42.x and 2.43.x trains.

Deprecated nodes are blocked at save time. Creating or updating a workflow that references a deprecated node type is now refused, surfacing the problem before a workflow fails at runtime. Both the 2.42.x and 2.43.x maintenance trains carry this change.

The Agent artifact appears while it's being built. A fix across both trains makes the agent editor display the artifact as the builder constructs it, rather than waiting until creation is complete — giving a clearer view of progress during agent development.

OpenAI-compatible providers work on self-hosted Assistant. Kimi K3 and other OpenAI-compatible models can now be selected on self-hosted n8n Assistant instances, after a regression that prevented them from loading correctly.

Agent builder prompts are eligible for caching. The system prompt for the agent builder is now kept static, making it a stable candidate for prompt caching and reducing latency on repeated invocations. The Instance AI skill sandbox is also prepared in the background, so it's ready before the first use.

Queue mode handles Chat Hub segments correctly. In queue-mode deployments, each Chat Hub segment is now processed exactly once, preventing duplicate handling.

Credential tests run in isolation. Concurrent credential tests no longer share a single node type registry, removing a race condition that could cause one test to corrupt another's context.

Google Gemini image model options now include Nano Banana 2.1 and 2 Lite on both active trains.