Branch Agent brings Git-style branching to LLM conversations

Short summary

Proposes Git-style branching for LLM conversations with O(1) storage via pointer references to parent context. Each branch can use different models, prompts, providers, and temperatures. Architecture uses Convex for atomic mutations and reactive queries, Agno Python SDK for agent configuration, and SSE streaming for real-time UI updates; includes merge logic with Judge Agent.

• •Apply Git branching model to LLM conversations—fork, branch, merge with different configs per branch
• •O(1) storage via parent pointers—no message duplication when creating branches
• •Full implementation with Convex backend, Agno agents, and streaming frontend

Generated with AI, which can make mistakes.