Agentic workflows normally postpone downstream work until a model or user selects a branch. DynBranch gives unresolved decision points stable coordinates, allowing likely branch subgraphs to execute during model decoding or user deliberation and completed results to be reused by later requests.
Reuse requires exact input matches and fresh external reads, while speculative results remain invisible until the selected branch authorizes them. A load-aware controller admits speculation only when its expected benefit exceeds its resource cost.
Across four workloads using Qwen3-32B on 4 H200 GPUs, DynBranch reduced mean latency by up to 32% versus each workload’s strongest prior system and by 46 to 66% versus no reuse. It operates at the model API boundary without modifying agent harnesses or inference engines.