n8n

The AI Builder in v2.33.0 gained several fixes that make agent-building sessions more predictable and less prone to stalling.

AI builder sessions no longer stall at approval gates. Workflow verification halts correctly at send-and-wait gates inside loops, so the builder no longer tries to advance past a human-review step. Channel configuration dropdowns for Linear and Telegram are now enabled during setup, and Slack credential labels are clearer about which access token is required. The builder also keeps its replies in the user's language during tool-heavy turns.

Security hardening landed across three release trains. S3 object paths are now signed with strict RFC 3986 encoding — a fix backported to the 2.31 and 2.32 trains — so pre-signed URLs no longer carry incorrectly percent-encoded characters. The credentials security audit no longer loads full execution data during its scan, reducing memory pressure and data exposure during audits. Security audit risk reporters are now imported with a file extension, fixing a module resolution issue.

The MCP Server Trigger node is more reliable. Execution data is saved only after a tool call completes, avoiding partial writes that could leave run records in an inconsistent state.

Scheduling and extensibility got targeted additions. The schedule trigger now accepts 5-field cron expressions alongside the existing 6-field format. Lifecycle external hooks receive the full workflow hook context, giving integrators access to richer state without custom workarounds.

The AI Assistant preserves sandbox provider selection across page loads, and Instance AI follow-up runs no longer loop indefinitely when they fail before the agent starts.