The AI agent stack has stabilized. Provider, model, instructions, tools, MCP, sub-agents, approvals, lifecycle hooks. Every framework ships some version of those parts now. When patterns settle like that, they're ready to collapse into config. So that's what we did.
In Lowdefy v5.3, a Lowdefy app can now run AI agents in just 30 lines of YAML 🤖
A connections entry for your provider. An agents entry with model, instructions, and a tool list. Any existing API endpoint becomes a tool by adding two fields: description and payloadSchema. Drop an AgentChat block on a page. That's the chatbot.
The interesting part is what comes for free. The same routine that fires when a user clicks a button can be called by the model mid-conversation. Same auth context. Same connections. Same operators. Same secrets. Same deployment.
Streaming, tool-call display, file attachments, tool approval cards, MCP servers, sub-agents, page state the agent can read and write, multi-provider routing through the Vercel AI Gateway, all in the same config shape.
Big credit to Machiel van der Walt for the design work on this one. The four-layer split, connection, agent, tool, block. Simplicity that feels inevitable 🏁
https://lnkd.in/dBR_hhNy