DecisionAct
An open-source task-focused agent framework that does not rely on an LLM as its reasoning core. Instead, it uses decision models (such as JEV and Cloudflare Clef) to make classification-style selections from predefined candidate actions and parameters, with execution carried out by Python functions defined by the developer.
Timeline (last 90 days)
DecisionAct is an open-source framework that uses decision models (such as JEV and Cloudflare Clef) rather than large language models to drive task-oriented Agents; developers define Python tools, input parameters, intents and reply templates, and the decision model is responsible for selecting actions and inputs.
DecisionAct v0.1.2 focuses on improving the reliability of multi-tool calls and the input parsing mechanism.
In DecisionAct v0.1.2, when the same tool is called repeatedly in succession, each repeated call re-parses the input and no longer reuses the previous parameters.
DecisionAct v0.1.2 supports chained execution, where the return value of a previous tool can serve as an input parameter for the next tool, for example first computing 4 plus 5 and then multiplying the result by 2.
DecisionAct v0.1.2 separates instructions, messages and completed tool calls within the execution context to reduce the possibility of the model confusing information from different sources.
In DecisionAct v0.1.2, when there is uncertainty in input selection, the model is given one chance to double-check; if it still cannot determine, it automatically asks the user a clarifying question.
DecisionAct v0.1.2 allows longer tool call sequences but sets a hard upper limit, and improves the ability to continue the flow after manual approval.
DecisionAct's input parsing supports three types of candidates: regular expression candidates, predefined classification options, and text candidates selected based on parameter descriptions.
DecisionAct currently supports only one intent per request, paired with multiple tool calls.
DecisionAct's installation command is pip install -U decisionact, and credentials for the chosen decision model provider must be prepared before use.
3 more timeline events
All Facts (13)
DecisionAct's model selection can still make mistakes, and users need to reserve fallback logic; the clarity of parameter descriptions directly determines input parsing accuracy, and in practice descriptions should be written as distinguishable discriminating features as much as possible, with a failure fallback set for each type of candidate.
50%UnverifiedDecisionAct is an open-source framework that uses decision models (such as JEV and Cloudflare Clef) rather than large language models to drive task-oriented Agents; developers define Python tools, input parameters, intents and reply templates, and the decision model is responsible for selecting actions and inputs.
50%UnverifiedIn DecisionAct v0.1.2, when the same tool is called repeatedly in succession, each repeated call re-parses the input and no longer reuses the previous parameters.
50%UnverifiedDecisionAct v0.1.2 supports chained execution, where the return value of a previous tool can serve as an input parameter for the next tool, for example first computing 4 plus 5 and then multiplying the result by 2.
50%UnverifiedDecisionAct v0.1.2 separates instructions, messages and completed tool calls within the execution context to reduce the possibility of the model confusing information from different sources.
50%UnverifiedIn DecisionAct v0.1.2, when there is uncertainty in input selection, the model is given one chance to double-check; if it still cannot determine, it automatically asks the user a clarifying question.
50%UnverifiedDecisionAct v0.1.2 allows longer tool call sequences but sets a hard upper limit, and improves the ability to continue the flow after manual approval.
50%UnverifiedDecisionAct's input parsing supports three types of candidates: regular expression candidates, predefined classification options, and text candidates selected based on parameter descriptions.
50%UnverifiedDecisionAct currently supports only one intent per request, paired with multiple tool calls.
50%UnverifiedDecisionAct's installation command is pip install -U decisionact, and credentials for the chosen decision model provider must be prepared before use.
50%UnverifiedDecisionAct's official documentation provides a quick start as well as three examples: a calculator, order support and a flight counter.
50%UnverifiedDecisionAct is positioned as a task-focused Agent rather than a general-purpose conversational assistant; when task boundaries are clear and tool interfaces are standardized, decision-model-driven approaches have advantages in reliability and interpretability, while open-ended tasks may still require LLM involvement.
50%UnverifiedDecisionAct v0.1.2 focuses on improving the reliability of multi-tool calls and the input parsing mechanism.
50%