Hi! I’m exploring a durable coding-agent architecture in BabyLovable and would love to compare approaches with others using AI SDK and Workflow.
The builder uses AI SDK v7 WorkflowAgent and Workflow DevKit for long-running agent turns. Each session has a remote Vercel Sandbox (Daytona is also supported) with a managed Next.js dev server and live preview. Supabase persists session state; Freestyle Git is the durable source of truth so an ephemeral sandbox can be reconstructed.
A recurring challenge is concurrency: preview requests, agent tools, and restarts can all request workspace readiness at once. We use a desired-state reconciler with a lease and version checks to avoid duplicate provisioning and stale state writes. For verification, the agent probes preview readiness and checks logs when compilation fails.
I’m curious how others handle three boundaries:
- When a durable run resumes after its sandbox has stopped, where do you place workspace reconciliation relative to agent/tool steps?
- How do you distinguish a healthy HTTP preview from a successful build or an actual UI regression?
- What strategies work well for keeping session streams and user-visible state consistent after reconnects?
The project is open source: https://github.com/1nFrastr/baby-lovable
Demo: https://baby-lovable.vercel.app/
Any feedback on the design or practical experience with these tradeoffs would be appreciated!