Google formalizes EGM: Design Long, Evaluate Short Memory

Google outlines a development framework where design occurs in a long session, followed by limited-context reviewers who evaluate only the draft and a rubric. This approach, structured in layers and phases, contrasts with traditional solicitation schemes and comes with operational limits.
Explicitly Addressed Costs, Tools, and Human Roles
The elephant-goldfish model involves additional costs and latency. Its implementation requires harnesses capable of orchestrating sub-agents. Furthermore, this method does not replace testing or human judgment.
A Workflow by Specification and Reviewers without Shared Memory
The model relies on a long design session, referred to as the "elephant," which drafts the specification. Then, review sessions called "goldfish" take over, focusing exclusively on this draft and a rubric, without consulting the original context, and this occurs before any development phase. The architecture revolves around layers dedicated to skills and commands, reviewer personas devoid of state memory, documentary standards, and evolving documentation that incorporates version management. Several steps structure the process: a brainstorming phase, draft writing, an organized review with prioritization, followed by a resolution and approval sequence integrated into an editing and validation cycle. After implementation, a drift report is produced if the code deviates from the specification. An example of delivery via webhook shows that the context of the "elephant" session can overlook gaps that limited-context reviewers detect. Google has published a repository for this model, accompanied by the slogan "Design with an elephant. Evaluate with a goldfish."
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Claimed Break from Common Solicitation Schemes
The model distinguishes itself from direct coding through single solicitation, self-evaluation in the same context, generic requests like "review this," and processes based on outdated documents.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.