An independent product exploration built on n8n: a pre-publish replay for understanding how changes to a governed agent affect its responses, approvals and actions.
The risky moment is before publish
An n8n agent can look valid on the canvas while a small change to its instructions, memory, model credential or tool policy materially changes what it says and does.
This prototype asks a focused question at that publishing moment: which representative conversations would end differently, and would the proposed agent gain any new external actions?
Replay behaviour, not just the graph
The published Relay workflow classifies intent and risk, retrieves approved policy, and then gives the agent a model, session-scoped memory and one constrained knowledge tool.
Each replay compares the current and proposed outcome. Differences are named as response changes, new tool calls, bypassed approval gates or failures, so a reviewer does not need to reverse-engineer the whole graph first.
Agentic execution needs deterministic guardrails
The agent can reason and propose an action, but deterministic n8n gates control what happens next. Missing policy is blocked, high-risk actions require human approval, and every demonstration write remains simulated.
The same published workflow now exposes two isolated, allowlisted release-control paths. One returns the real n8n execution ID, support fixtures and node trace; the other persists a pending approval in an n8n Data Table. Neither path calls production data or changes an external system.
Test the support agent before it ships.
Replay realistic customer conversations, inspect the agent’s trajectory and turn every failure into a reviewable quality signal.
CRM action
External write enabled, approval removed
Approval required
The draft bypasses the human gate for a high-risk action. Relay should not publish this change without review.