Paper codifies end-to-end methodology for building custom AI agents
A paper by Marc Alier Forment, Juanan Pereira, and Francisco José García-Peñalvo formalizes "Agents All the Way Down," a framework-free methodology for building production-ready custom AI agents from two preconditions and three repeating practices.
Score breakdown
The paper fills a documented gap by writing down, for the first time in a consolidated form, the end-to-end practice for building production custom AI agents — knowledge the authors note has previously existed only in informal sources like podcasts, blogs, and leaked system prompts.
- 01The paper formalizes a methodology called 'Agents All the Way Down' for building custom AI agents end to end.
- 02Custom agents are defined as agents with their own application, data, tools, security boundaries, brand, and audit trail.
- 03The methodology comprises two preconditions (P1: Substrate; P2: Building Blocks) and three repeating practices (P3: prototype, P4: Turtle pattern CLI harvest, P5: agent-tests-agent).
The paper introduces "Agents All the Way Down," a methodology for building custom AI agents — defined as agents that live inside their own application, operate on their own data and tools, enforce their own security boundaries, and carry their own brand and audit trail. The authors distinguish custom agents from general-purpose agents by fit rather than capability: each is built for a single job by the engineer who will maintain it. The paper's central motivation is that while the component pieces (function-calling APIs, the Model Context Protocol, code agents) are widely available, the practice of chaining them into a production agent has been scattered across informal sources with no consolidated written methodology.
The methodology is structured as two preconditions crossed once and maintained, followed by three practices repeated throughout the agent's lifecycle.
The methodology is structured as two preconditions crossed once and maintained, followed by three practices repeated throughout the agent's lifecycle. Precondition P1 (Substrate) frames the LLM as a software component organized around tools, system prompts, messages, and prompt-caching. Precondition P2 (Building Blocks) covers function calling, MCP, CLI orchestration, the liteshell pattern, the agent loop, skills, characters, hooks, and scaffolding. The three practices are: P3, prototype with a general-purpose agent; P4, harvest and fold the result into a CLI using the "Turtle pattern"; and P5, agent-tests-agent, in which a general-purpose agent drives the custom agent through behavioral scenarios as a complement — not a replacement — to classical testing. The working loop cycles through P3 → P4 → P5 and back. A corollary the authors derive is that multi-agent orchestration is simply CLI composition.
The methodology is framework-free and language-agnostic by construction. It was distilled from the AAC, a custom agent built for the open-source LAMB platform, developed in approximately ten days by a single developer working with an AI pair-programmer and already in production. The authors present it as a transferable practice applicable across languages and frameworks.
Key facts
- 01The paper formalizes a methodology called 'Agents All the Way Down' for building custom AI agents end to end.
- 02Custom agents are defined as agents with their own application, data, tools, security boundaries, brand, and audit trail.
- 03The methodology comprises two preconditions (P1: Substrate; P2: Building Blocks) and three repeating practices (P3: prototype, P4: Turtle pattern CLI harvest, P5: agent-tests-agent).
- 04P5 uses a general-purpose agent to drive the custom agent through behavioral scenarios, complementing but not replacing classical testing.
- 05A key corollary: multi-agent orchestration is equivalent to CLI composition.
- 06The methodology is framework-free and language-agnostic by construction.
- 07It was distilled from the AAC agent for the open-source LAMB platform, built in ~10 days by one developer with an AI pair-programmer and already in production.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 11, 2026 · 08:34 UTC. How this works →