Baton Orchestration Framework

• 

  baton

Inspiration

The project started after I discovered that the "ultracode" effort level in Fable 5 weren't actually using the highest available reasoning levels. Instead, it primarily added agent delegation while running the orchestrator at lower reasoning settings ("xhigh" in Fable 5 instead of "max").

That felt wasteful. Those modes consume significantly more tokens, yet they don't combine delegation with the models' strongest reasoning capabilities. I wanted to build an orchestration framework that does both: use the highest reasoning level possible for the orchestrator while still benefiting from specialized worker agents.

What it does

Baton is an orchestration framework for coding agents.

A single orchestrator runs at the highest reasoning level available (if chosen) and delegates work to specialized workers with three difficulty levels: Easy, Medium, and Hard.

Each worker can use any harness, provider, model, and reasoning level, making the framework portable instead of being tied to a specific AI platform.

To improve output quality, Baton also performs context management. Before delegating work, the orchestrator inserts the most important parts of the user's request at both the beginning and the end of each worker's context. This is inspired by the "Lost in the Middle" research, which showed that language models pay less attention to information placed in the middle of long contexts.

The project is fully open source on GitHub and includes a ready-to-use installation, along with two prompts: one that generates the framework and another that verifies the generated result. This makes Baton adaptable to future models, allowing stronger models to regenerate and improve the framework while remaining compatible with any supported harness.

How we built it

Baton was built entirely through vibe coding.

I started by creating a small prototype using Fable 5 and GPT-5.6 Sol. Once the initial version worked, I began using Baton itself to improve Baton. The framework went through multiple iterations, with each version helping develop the next one.

The result is a self-improving development workflow where the framework became one of its own primary development tools.

Challenges we ran into

Debugging on Sol consumed the majority of the development tokens. Getting orchestration, delegation, and context management to work reliably across different configurations required many iterations.

Another challenge was deciding which features belonged in the framework. I wanted Baton to stay as lean as possible while still including functionality that genuinely improves code quality and reduces token consumption. Finding that balance took several rounds of refinement.

Accomplishments that we're proud of

• Built a provider-agnostic orchestration framework that works with any harness, provider, model, and reasoning level.
• Implemented intelligent context management inspired by the Lost in the Middle research.
• Created a framework that can generate and verify itself using included prompts.
• Successfully used Baton to iteratively improve Baton throughout development.
• Released the entire project as open source with ready-to-use installation and documentation.

What we learned

Building Baton reinforced that orchestration alone can significantly improve coding quality, but combining orchestration with the strongest available reasoning level can produce even better results.

We also learned how much thoughtful context management matters. Small changes to how information is presented to language models can noticeably improve output quality without simply increasing context size or token usage.

What's next for Baton

• benchmarking different models and reasoning levels vs "ultracode" and "ultra" (high priority),
• more orchestration strategies,
• smarter task difficulty evaluation,
• improved context management techniques,
• and continuously regenerating the framework with newer, more capable models as they become available.

Built With

• baton
• claude-code
• fable-5
• gpt-5.6
• hermes
• reasearch-papers
• skills
• sol