DecisionLoop
A Python agent framework that splits agent work into a language layer and a decision layer. The decision layer is handled by a dedicated small decision model that performs intent recognition, tool selection, parameter generation and flow control, rather than having the LLM reason from scratch on every turn, with the aim of reducing the latency and cost of tool invocation.
Timeline (last 90 days)
DecisionLoop is a Python framework that decouples an Agent's decision layer (intent recognition, tool selection, parameter generation, process control) from the language layer, with the decision layer handled by a small decision model and the language layer handled by an LLM for natural language understanding and generation.
DecisionLoop was validated on three business scenarios (order customer service, flight help desk, and banking help desk), which share the common traits of relatively clear processes and a limited tool set while also needing to handle the diversity of user expression.
Early test results show that DecisionLoop outperforms an LLM Agent built on LangChain with a 12B model in tool-calling accuracy, while also having advantages in cost and latency.
The author himself calls these preliminary results preliminary observations rather than formal benchmark tests, and the test scenarios, sample size, and evaluation criteria have not been made public.
The author's next step is to build a multi-Agent architecture on top of DecisionLoop, with a routing layer dispatching requests to domain expert Agents; routing itself is also a decision problem and can be handled by a small model.
All Facts (5)
DecisionLoop is a Python framework that decouples an Agent's decision layer (intent recognition, tool selection, parameter generation, process control) from the language layer, with the decision layer handled by a small decision model and the language layer handled by an LLM for natural language understanding and generation.
50%UnverifiedDecisionLoop was validated on three business scenarios (order customer service, flight help desk, and banking help desk), which share the common traits of relatively clear processes and a limited tool set while also needing to handle the diversity of user expression.
50%UnverifiedEarly test results show that DecisionLoop outperforms an LLM Agent built on LangChain with a 12B model in tool-calling accuracy, while also having advantages in cost and latency.
50%UnverifiedThe author himself calls these preliminary results preliminary observations rather than formal benchmark tests, and the test scenarios, sample size, and evaluation criteria have not been made public.
50%UnverifiedThe author's next step is to build a multi-Agent architecture on top of DecisionLoop, with a routing layer dispatching requests to domain expert Agents; routing itself is also a decision problem and can be handled by a small model.
50%