Dom Sipowicz, Solutions Architect at Vercel, shared a practical framework for thinking about agents as a new kind of software: systems that automate workflows that were previously too complex or too expensive to automate. Rather than focusing on fully autonomous systems as a goal in itself, he argued for a more useful middle ground where teams build agents around real business workflows and production constraints. The talk traced that shift from simple single-model AI apps, to chatbots with memory and tools, to asynchronous background agents that can be triggered from systems like email, Slack, CRMs, or scheduled jobs. From that perspective, the real challenge is not defining agents in abstract terms, but designing them so they are economically viable, fast to build, observable, and secure.

He then mapped that architecture to the Vercel stack. Vercel AI SDK provides the core primitives for model access and tool use while keeping provider switching simple. Vercel Workflows handles the long-running parts of agent systems, including retries, persistent state, human-in-the-loop pauses, and observability, without forcing teams to assemble queues, caches, and recovery logic by hand. Vercel Sandbox extends beyond code generation into a broader runtime for CLI-driven agents that can operate on files, data, and Linux tooling inside an isolated environment. The talk also covered AI Gateway for centralized model access and cost control, BotID and Firewall for protecting agentic endpoints, and the broader AI Cloud idea of mapping framework code to the infrastructure primitives needed to run it.

What you’ll walk away with:

• How to think about agents as workflow automation rather than just chat interfaces
• Why the most useful agent use cases sit between traditional software and fully autonomous systems
• How Vercel Workflows supports retries, state, human approval steps, and observability in long-running AI systems
• Why Sandbox can power CLI-driven agents and file-based tasks, not just code generation
• How AI Gateway, BotID, Firewall, and observability fit into a production architecture for agentic systems

Watch the recording to see how teams can design agentic systems that are durable, observable, and ready for production workflows.