New 'Agents All the Way Down' methodology builds production AI agents in 10 days

A new paper from Marc Alier Forment and colleagues at Universitat de Lleida presents 'Agents All the Way Down,' a practical methodology for building custom AI agents from scratch to production. Unlike general-purpose AI assistants, custom agents live inside their own application, talk to their own data and tools, enforce security boundaries, and carry their own brand and audit trail. The authors argue that while the components — function-calling APIs, the Model Context Protocol (MCP), code agents — are scattered across podcasts and blogs, no published practice chains them together. Their methodology fills that gap with a reproducible, framework-free approach.

The methodology consists of two permanent preconditions and three iterative practices. Precondition 1 (Substrate) treats the LLM as a software component via tools, system prompts, and message caching. Precondition 2 (Building Blocks) defines function calling, MCP, CLI orchestration, the liteshell pattern, agent loops, skills, characters, hooks, and scaffolding. The three practices are: P3 — prototype with a general-purpose agent; P4 — harvest, fold, and ship the result as a CLI (the 'Turtle pattern'); and P5 — agent-tests-agent, where a general-purpose agent drives the custom agent through behavioral scenarios. The working loop cycles through P3→P4→P5 and back. A natural corollary: multi-agent orchestration is simply CLI composition. The authors validated the methodology by building AAC, a custom agent for the open-source LAMB platform, in about ten days by one developer with an AI pair programmer. The result is a transferable practice independent of any language or framework.