AgentFlow
At a glance
- Shipsy’s visual workflow and agent builder — think n8n for logistics AI.
- The platformization layer that lets customers and CS compose agents without engineering.
- Ships with an AI Copilot that builds workflows from plain English.
- Knows 1,862 Shipsy fields and 65+ workflow blocks.
- The single highest-leverage tool for new joiners to learn in week 1.
Why this matters
AgentFlow is what makes Shipsy a platform, not a professional-services company. Without it, every new agent or automation requires an engineering project. With it, a CS person can ship a working workflow in hours. This is the tool that scales customer value without scaling headcount.
How AgentFlow Copilot works
The Copilot lets users describe what they want in plain English. It translates the request into a workflow by:
- Mapping the description to the right workflow blocks (from 65+ available)
- Selecting the correct fields (from 1,862 Shipsy-native fields)
- Generating the YAML configuration behind the scenes
- Presenting a visual workflow the user can review, edit, and activate
Powered by Claude (Sonnet 4.6) via Managed Agents.
Example
“When an order has 2 failed delivery attempts, send a WhatsApp to the customer with the driver’s number and a reschedule link.”
Copilot generates: trigger block (order status = 2 failed attempts) → data fetch block (customer phone, driver number) → WhatsApp block (template message with reschedule link).
Permissions model
AgentFlow uses a granular permissions framework:
| Permission | What it controls |
|---|---|
AGENTFLOW_WORKFLOW_CREATE | Create new workflows |
AGENTFLOW_WORKFLOW_UPDATE | Edit existing workflows |
AGENTFLOW_WORKFLOW_DELETE | Remove workflows |
AGENTFLOW_WORKFLOW_TOGGLE | Enable/disable workflows |
AGENTFLOW_WORKFLOW_VIEW_LOGS | View execution logs |
Permissions are set at the org level. If a customer reports “I can’t see/create workflows”, check these first.
Triggers
Workflows can be triggered by:
- Webhook — external systems call an AgentFlow endpoint with a payload
- Event — internal Shipsy events (order created, status changed, SLA breached)
- Schedule — cron-based recurring execution
Webhook payloads must match the expected schema — malformed payloads are a common support issue.
Execution logs
All workflow executions are stored in Elasticsearch with a retention window of approximately 1-2 months. To debug a failed execution:
- Open the workflow in AgentFlow
- Click “Execution logs”
- Find the failed run by timestamp
- Inspect each block’s input/output to find where it broke
Common support patterns
From the #agentflow-support channel — issues new CS joiners hit most:
| Issue | Root cause | Fix |
|---|---|---|
| ”Workflow won’t save” | Missing AGENTFLOW_WORKFLOW_CREATE permission | Check org permissions |
| ”Webhook not firing” | Payload format mismatch | Validate JSON against expected schema |
| ”Can’t see execution logs” | Missing VIEW_LOGS permission | Grant at org level |
| ”Workflow runs but does nothing” | Block misconfiguration (wrong field mapping) | Inspect block config in edit mode |
AgentFlow vs AgentFleet Builder
See the comparison table on the AgentFleet page. Short version: AgentFlow for standard patterns and customer self-serve; AgentFleet Builder for bespoke/deep agents.
Try it yourself
- Get access to the demo org (ask your manager or check the AgentFleet page for URLs).
- Open AgentFlow in the demo org.
- Try the Copilot: describe a simple workflow in English and see what it generates.
- Manually build a 3-block workflow: trigger → data fetch → notification.
- Check the execution log after it runs.
Sources
- Slack: #tms-product-engineering-team — Sarvagnya Jatti (17 Apr 2026): AgentFlow Copilot launch
- Slack: #agentflow-support — common patterns
- Agentic Capabilities — Master deck
Changelog
- 26 May 2026: Enriched with Copilot details, permissions model, support patterns, and execution log guidance.