Skip to content
NewsProduct launch

n8n 2.45 extends agent budget stops to sub-agents and logs which model answered

· by Pondero Newsdesk

The short version

n8n's version 2.45 release closes a spend-tracking gap by applying budget alerts and stops to sub-agent runs, not just the top-level agent, and now records which model actually generated an LLM node's response.

n8n 2.45 extends agent budget stops to sub-agents and logs which model answered

A spend cap that only watched the top-level agent used to let a workflow's sub-agents keep running up API costs unseen. n8n's version 2.45 release, tagged October 20, 2026, closes that gap and adds a way to see which model actually generated an answer when a node is configured to pick one dynamically, per the release notes on GitHub.

What shipped

Budget alerts and automatic stops, previously enforced only at the top-level agent node, now apply to sub-agent runs as well, so a workflow that spins up nested agents cannot quietly exceed its spend limit through a child process the cap never watched. The AI execution log now surfaces the name of the model that actually generated an LLM sub-node's response rather than only the configured default, useful for workflows that route between models dynamically. The Text Classifier node gained an opt-in confidence score on its output, and n8n added automated migrations plus notifications, sent to individual and bulk workflow owners, when a removed node needs replacing. Separately, the Monday.com node's V2 version became the default with full compatibility for workflows still built on V1, and the Kafka node's V2 version became the default for new instances, per the same release notes. n8n also turned on node and credential type policies by default and now shows the specific policy violation reason when an agent's tool call gets refused, tightening governance defaults for teams running agents against systems they would rather keep locked down.

The blind spot this closes for unattended workflows

Teams running n8n agents without a human watching every execution have had no way to confirm a nested sub-agent stayed inside budget once the parent agent handed off a task. That gap mattered most for multi-step agent chains, the kind n8n markets for support triage or research workflows, where a single run can spawn several sub-agent calls before finishing. The model-attribution logging addresses a separate but related trust problem: a node set up to fall back between models, say a primary and a cheaper backup, previously left no record of which one actually answered, making it hard to audit cost or quality after the fact.

What to watch next

Watch whether n8n adds a dashboard or alerting layer on top of the new budget data rather than leaving it in the execution log alone, and whether the automated node-migration notifications reduce the manual cleanup work that follows a removed-node update.

Sources