OpenAI's Agent Builder shuts down November 30, 2026 — about fourteen months after it launched. Here's how to rebuild your workflows somewhere built to last.
An Agent Builder workflow is a handful of agents, some branching logic, and a set of tools. That maps onto LiveGraph directly — the difference is what happens after you press run. On LiveGraph the canvas stays live: every hop is visible while it's executing, a run parks at a human approval gate if you want one, and dragging an edge mid-run steers the run you're watching.
| Agent Builder | LiveGraph |
|---|---|
| Agent nodes (instructions + model) | Worker/router nodes — role, prompt, and model per node; any provider, not just OpenAI |
| If/else branches and state | explicit edges, or auto edges where the model picks the next specialist |
| Human approval nodes | Approval gates — the run parks at /approvals with the pending payload visible |
| MCP server tools | Per-node MCP servers with an exact-name tool allowlist; tokens encrypted at rest |
| Guardrails | A reviewer node — its verdict is a routing decision, not a silent block |
| Preview/Test chats | Runs start live from the canvas — you watch them hop by hop and can redirect |
We don't offer a one-click import of Agent Builder exports — we'd rather say that than half-convert your workflow. In practice a rebuild is fast: most workflows are a few agents and a few branches, and describing a node in words fills in its role, prompt, and model for you. If you want to see the mechanics first, the live demo runs a real multi-agent graph in your browser with no signup and no API key — including the mid-run edge drag.
http_request endpoint, scoped per node.Agent Builder is the second hosted agent canvas to close this year — Flowise reached end of life in August (see our Flowise migration guide for its importer). LiveGraph's answer to "will this exist in 18 months" is a flat team price instead of per-run metering, provider keys that stay yours (or hosted ones if you'd rather not manage them), and a canvas that stays editable while a run is executing — no black box to abandon.
Create an account, describe your first agent, and rebuild your simplest workflow — you'll know within one run whether the canvas earns the rest.