Bonginkan Releases Research OSS 'Fairy Tale' to Share Successful AI Agent Workflows

Bonginkan Co., Ltd. (Headquarters: 1-2-2 Yurakucho, Chiyoda-ku, Tokyo; Representative Director: Hiroki Tsubouchi) announces the release of 'Fairy Tale,' an open-source software (OSS) that systematically organizes and provides reproducible workflow enhancement skills based on public information and user reports regarding AI agents.

Fairy Tale does not modify AI models themselves. It is a research project that studies reports and case studies of Fable-class and Mythos-class AI agents, extracting and organizing reusable workflows and best practices from them.

Hosted on GitHub under the Apache License 2.0, Fairy Tale is provided as skills and plugins compatible with agent-enabled development environments such as Claude Code and Codex.

## Background

In recent years, numerous successful cases have been reported where AI agents have achieved high performance in software development, research tasks, and document creation.

In particular, high-performance agents referred to as Fable-class or Mythos-class have attracted significant attention.

However, many of their capabilities depend on specific services or access environments. Changes in usage conditions or access restrictions may prevent access to excellent workflows and operational insights.

Moreover, it remains unclear whether superior results stem from model performance, workflow design, or feedback loops, making reproduction and verification difficult.

Fairy Tale was conceived from the idea: 'Instead of consuming excellent agent outcomes as myths, can we preserve the reproducible procedures behind them?'

The project name 'Fairy Tale' draws inspiration from Hans Christian Andersen's fairy tale 'The Nightingale.'

In the story, the living nightingale's song is portrayed as having more intrinsic value than the jewel-encrusted mechanical bird.

Similarly, Fairy Tale focuses not on the AI model as a 'machine' itself, but on the reproducible knowledge and workflows behind it.

## Features of Fairy Tale

■fairytale

https://fairytale.run/

### Research Approach Limited to Public Information

Fairy Tale targets only publicly available official information and user reports for research. It does not attempt to bypass access controls, break model protection mechanisms, or obtain non-public data.

The goal is to analyze published cases and organize them into reproducible procedures.

### Compatibility with Claude Code and Codex

Fairy Tale is provided in the following formats:

- General agent skills
- Skills for Claude Code
- Skills for Codex
- Plugins for Claude Code
- Plugins for Codex

This enables the same workflows to be utilized across different agent environments.

### Continuous Improvement via Self-Feedback

Fairy Tale implements a 'self-feedback mechanism' that improves performance by refining workflows, without retraining the model.

This system forms a closed loop consisting of:

- Executing tasks
- Classifying failure patterns
- Converting them into general improvement rules
- Re-evaluating under the same conditions
- Automatically removing ineffective rules

For example, it classifies and accumulates reusable improvement rules for cases such as:

- Missing only one item to pass
- Failures due to insufficient evidence
- Failures from unnecessary decomposition
- Format mismatches

Additionally, rules that show no improvement effect or are contradictory are automatically pruned, preventing knowledge bloat.

## Validation Results at Research Stage

Fairy Tale has validated the effectiveness of workflow improvements across multiple domains, including biology, law, and coding. In publicly shared local evaluations:

- On BioMysteryBench-preview, it achieved 80.0% compared to 60.0% for GPT-5.5 alone
- On a Harvey LAB-compatible legal benchmark, it achieved 11.0% compared to 2.1% for GPT-5.5 alone

Results close to reported values of publicly known Fable/Mythos-class agents have also been observed in some benchmarks.

Furthermore, in legal evaluations applying the self-feedback mechanism:

- Full criteria pass rate improved from 0.0% to 20.0%
- Criterion Pass rate improved from 83.21% to 90.61%
- One-Miss failures decreased from 10 to 5

These are local validation results at the research stage and do not represent official rankings or vendor-published values.

https://fairytale.run/

https://fairytale.run/

## Research Process Emphasizing Reproducibility

Fairy Tale does not treat published cases or user reports as facts by default. Instead, it emphasizes independent re-verification.

Research notes, verification plans, evaluation results, and best practices are all published within the repository, enabling third parties to track and verify accumulated knowledge.

## Developer Comments

Fairy Tale Developer: The Pioneer

'There are many reports about powerful AI agents. However, mechanisms for sharing the workflows behind these results in a reusable manner have been insufficient.

We developed Fairy Tale from the idea: 'If it doesn't exist, why not create it?'

We believe it is important not to entrust the development of intelligence solely to closed environments, but to preserve the freedom to verify, reproduce, and improve publicly shared knowledge.

Fairy Tale is a project to document the 'reproducible song' behind excellent agent outcomes, rather than treating them as myths.

Our motto is 'Let's fable the model.''

## Future Outlook

Bonginkan Co., Ltd. will continue advancing research and sharing reproducible workflows and best practices for AI agent utilization through Fairy Tale.

Additionally, through collaboration with the OSS community, we aim to create an environment where more developers can verify and reuse knowledge.

■fairytale

https://fairytale.run/

■GitHub

Fairy Tale GitHub Repository
https://github.com/bonginkan/fairy_tale

■About Bonginkan Co., Ltd.

Bonginkan Co., Ltd. is a company engaged in solution development and R&D utilizing AI technology. We work in fields such as AI agents, AI telephony, and AI avatars, aiming to promote practical AI adoption.

https://bonginkan.ai/