Workflow agent “outside-in” build guide: dual-loop design and tool selection

O’Reilly Radar published a July 20, 2026 session recap with AI consultant John Berryman on how to build a workflow agent using a simple “two-loop” pattern. In Berryman’s framing, an agent is a large language model wrapped in two loops. The outer loop handles back-and-forth messaging with the user (like chatbots). The inner loop lets the model choose and run tools, turning an assistant into an agent. He argues the structure has stayed stable since 2023, while tools and instructions have improved enough to support real products. He recommends an “outside in” development approach: build traditional software first (UI, data model, non-AI components), then define the AI input/output requirements. Next, fake the agent with a stub to verify integration. Replace the stub with a minimal real agent that uses Pydantic’s agent plus an AI reviewer, and enforce structured output with validation. For tools, he suggests starting small—read, write, edit, and shell access—so the agent can extend capabilities when needed (e.g., running curl for research). Instead of coding a state machine, write the agent’s context, decision criteria, and step-by-step workflow in natural language, and embed checklists so the agent confirms completion or fails fast. Berryman also predicts a shift from purpose-built apps toward more general agents that work across tools and interfaces on a user’s behalf.