Ornith 1: Self-Improving Agentic Coding Models

The Breakthrough

Ornith 1 models utilize a reinforcement learning technique (GRPO) to train the model to generate both a solution and a task-specific harness—including memory, retries, and error handling—in a single loop, allowing the model to self-scaffold for specific coding tasks.

What Actually Worked

• The training process uses Group Relative Policy Optimization (GRPO) to reward both the generated solution rollouts and the custom harness structure simultaneously.
• To prevent reward hacking, the researchers implemented three layers of defense: locked boundaries for environment tools, deterministic monitoring to flag unauthorized file access, and a frozen judge model to verify outputs even when they pass initial tests.
• The model dynamically constructs a harness on the fly based on task complexity, which is discarded after the task is completed.
• Testing the 9B parameter version against the Qwen 3.5 9B base model showed comparable accuracy but significantly higher efficiency, with costs reduced by approximately 3× on average and up to 20× on specific tasks.

Context

Agentic coding models often rely on human-written harnesses to manage memory and error handling. Ornith 1 shifts this responsibility to the model itself. While the 9B model demonstrates efficiency gains, the author notes that long-horizon reasoning—such as maintaining honesty when presented with false user claims—appears to require the larger 35B+ parameter models to function reliably.

Content References

• tool: Terminal Bench, mentioned
• tool: Qwen 3.5, mentioned
• tool: Gemma 2, mentioned
• tool: Ollama, mentioned
• tool: Anthropic Dynamic Workflow, mentioned