Peter Wielander is a software engineer at Vercel and a founding contributor to the open-source Workflow SDK, which gives TypeScript applications and AI agents persistent execution, automatic recovery, and operational visibility. He focuses on making agents dependable when their work outlasts individual requests, deployments, or serverless instances.
Before Vercel, Wielander was a staff software engineer and technical lead for emerging products at Minerva Project, developing technology for active learning in higher education. He also created RealPhabet, a tool for practicing unfamiliar writing systems.
His AI Engineer Code workshop converted a coding agent built with AI SDK, Next.js, and Vercel Sandbox into a recoverable workflow. His approach turns ordinary TypeScript functions into deterministic orchestration and individually persisted execution steps, making model requests and tool calls observable, retryable, and resumable.
- Durable execution without bespoke infrastructure: Workflow SDK provides persisted state, retries, and inspection without requiring each application team to assemble separate queues, databases, and recovery mechanisms.
- Resumable streams and human approval: Clients can reconnect to interrupted agent output, while webhook-based approvals and scheduled tasks suspend workflows without continuously consuming compute.
- Safe retries and explicit system boundaries: Side effects belong in execution steps; developers remain responsible for idempotency and permissions. Workflow durability does not preserve an external sandbox’s state or provide security isolation.
- Deployment isolation and workflow versioning: Wielander’s workflow-versioning proposal keeps existing runs pinned to their originating deployment and treats upgrades as deliberate compatibility decisions.
- Framework-portable developer tooling: He coauthored Workflow SDK support for TanStack Start, expanding durable execution across the TypeScript ecosystem.
As AI-generated code becomes cheaper, Wielander argues that durable, observable production infrastructure becomes more important: generating an application is easier than ensuring its agents remain understandable and reliable.
