AI Agent Runtime 'Airuncode' Released, Supporting Multi-Model Debate and Self-Healing

Developer Gustavo Arretureta has released local AI agent runtime 'Airuncode'. This tool supports Windows, macOS, and Linux, and adopts a model-agnostic structure that allows users to enter their own API keys to freely switch between cloud models or local models (such as LM Studio and Ollama).

Multi-Agent Debate and Code Duplication Prevention

Airuncode is designed so that each agent can use completely different models simultaneously during parallel execution. This implements a debate loop to compare solutions from different perspectives. To prevent code duplication and context drift, it generates a global symbol map and AST index of the repository before debate or editing. If Agent A proposes code that duplicates an existing implementation, Agent B detects this and suggests reusing the existing implementation instead of introducing duplication. However, the original text acknowledges the limitation that AST-based structural matching cannot catch semantic duplications such as 'fetch wrapper vs retry loop'.

Structural Self-Healing Mechanism

When tests fail, it applies Stack Trace Primacy to resolve the root cause rather than simply making the test pass (Green Test). The runtime forces the stack to be traced up to the originating frame where the actual cause lies, rather than the failure point (top frame). Additionally, Scope Locking strictly limits the scope of modifications, allowing explicit privilege escalation only when actual changes are needed in another layer. After self-healing, it goes through an architecture review step to verify whether there is layer pollution or increased coupling.