Bonginkan Releases “Fairy Tale,” an Open-Source Research Project for Reproducible AI Agent Workflows
Reproducing Fable-Class Benchmark Results with Existing Models Through Workflow Enhancements
Bonginkan Co., Ltd. (Head Office: 1-2-2 Yurakucho, Chiyoda-ku, Tokyo, Japan; CEO: Hiroki Tsubouchi) today announced the release of Fairy Tale, an open-source software (OSS) project that studies public AI agent reports and user observations, transforming them into reproducible workflow-enhancement skills and reusable best practices.
Fairy Tale does not modify AI models themselves. Instead, it focuses on publicly reported capabilities and case studies surrounding Fable-class and Mythos-class AI agents, extracting reusable workflows, operational patterns, and best practices that can be applied across agent environments.
The project is released under the Apache License 2.0 and is available on GitHub as skills and plugins compatible with agent-oriented development environments such as Claude Code and Codex.
Background
Recent advances in AI agents have produced remarkable results in software engineering, research, analysis, and document generation.
Among these developments, high-performing systems often described as Fable-class or Mythos-class agents have attracted significant attention.
However, many of these capabilities depend on specific services, access conditions, or proprietary environments. As access policies change, valuable workflow knowledge and operational techniques may become difficult or impossible to study and reproduce.
Furthermore, it remains unclear how much of a reported success comes from:
-
Model capability
-
Workflow design
-
Feedback loops and evaluation processes
As a result, many reported achievements are difficult to verify, reproduce, or build upon.
Fairy Tale was created from a simple idea:
Rather than consuming impressive agent results as mythology, can we preserve the repeatable processes behind them?
The project name “Fairy Tale” is inspired by Hans Christian Andersen’s classic story The Nightingale.
In the tale, a living nightingale ultimately proves more valuable than a jeweled mechanical bird.
Likewise, Fairy Tale focuses not on the AI model as a machine, but on the reproducible knowledge, workflows, and practices behind successful outcomes.
Key Features of Fairy Tale
■Official Website
Research Based Exclusively on Public Information
Fairy Tale studies only publicly available official information and public user reports.
The project does not attempt to bypass access controls, circumvent model safeguards, or obtain non-public information.
Its goal is to analyze publicly documented outcomes and organize them into reproducible workflows.
Available for Claude Code and Codex
Fairy Tale is distributed as:
-
Generic agent skills
-
Claude Code skills
-
Codex skills
-
Claude Code plugins
-
Codex plugins
This allows the same workflow improvements to be reused across different agent environments.
Continuous Improvement Through Self-Feedback
Fairy Tale includes a self-feedback mechanism designed to improve performance through workflow optimization rather than model retraining.
The system operates as a closed-loop process:
-
Execute a task
-
Classify failure patterns
-
Convert observations into reusable improvement rules
-
Re-evaluate under identical conditions
-
Automatically prune ineffective rules
For example, Fairy Tale identifies and categorizes failures such as:
-
One-miss cases that narrowly fail completion criteria
-
Insufficient evidence or justification
-
Excessive task decomposition
-
Formatting inconsistencies
These observations are converted into reusable workflow rules.
Rules that fail to demonstrate measurable improvement—or that conflict with other rules—are automatically pruned, preventing knowledge accumulation from becoming noisy or contradictory.
Research Results
Fairy Tale currently evaluates workflow-improvement techniques across multiple domains, including biology, legal reasoning, and software engineering.
Published local validation results include:
-
BioMysteryBench-preview: GPT-5.5 baseline 60.0% → Fairy Tale 80.0%
-
Harvey LAB-compatible legal benchmark: GPT-5.5 baseline 2.1% → Fairy Tale 11.0%
In several evaluations, Fairy Tale also demonstrated performance approaching publicly reported Fable/Mythos benchmark results.
Applying Fairy Tale’s self-feedback mechanism to legal evaluations produced additional improvements:
-
All-pass rate: 0.0% → 20.0%
-
Criterion pass rate: 83.21% → 90.61%
-
One-miss failures: 10 → 5
■Fairy Tale
These measurements are local research validations and should not be interpreted as official benchmark submissions or vendor-published leaderboard scores.
A Research Process Focused on Reproducibility
Fairy Tale does not treat public reports as established facts.
Instead, the project emphasizes independent reproduction and verification.
Research notes, evaluation plans, benchmark results, and best-practice documentation are published within the repository, enabling third parties to inspect, reproduce, and validate findings.
Developer Comment
The Pioneer, Creator of Fairy Tale
“There are countless reports describing powerful AI agents. Yet there has been no widely adopted framework for sharing the workflows behind those results in a reusable form.
We built Fairy Tale from a simple belief: if it doesn't exist, why not create it?
We believe the advancement of intelligence should not depend solely on closed environments.
Developers should have the freedom to inspect, reproduce, and improve upon publicly available knowledge.
Fairy Tale is not about turning successful agents into myths.
It is about recording the reproducible song behind the performance.
Our motto is:
‘Let's fable the model.’”
Future Outlook
Through Fairy Tale, Bonginkan will continue researching and sharing reproducible workflows and best practices for AI agent development and deployment.
The company also aims to collaborate with the OSS community to create an environment where developers can verify, reuse, and improve upon shared knowledge.
■Fairy Tale
■GitHub Repository
https://github.com/bonginkan/fairy_tale
About Bonginkan Co., Ltd.
Bonginkan Co., Ltd. develops AI-powered solutions and conducts research and development in applied artificial intelligence.
The company works across areas including AI agents, AI-powered telephony, and AI avatars, with the goal of expanding practical AI adoption and creating accessible tools for real-world use.
すべての画像
