A New Language Makes AI Reasoning More Accessible

Programming intelligent systems that can adapt to uncertain environments has long been a , hindered by complex symbolic languages that are difficult for newcomers to master. Non-Axiomatic Reasoning Systems (NARS) offer a framework for building such adaptive agents, but their standard input language, Narsese, uses dense notation with overloaded punctuation and implicit conventions, making programs hard to read, write, and maintain. This barrier has limited NARS adoption outside specialist circles, slowing experimentation and practical application. DriftScript, a new domain-specific language, aims to change that by providing a more accessible way to program these reasoning agents, enabling broader use in autonomous systems that must operate with insufficient knowledge and resources.

Researchers have developed DriftScript as a Lisp-like language that compiles directly to Narsese, covering the major sentence and term forms used in Non-Axiomatic Logic (NAL) levels 1 through 8. This includes constructs for inheritance, temporal implication, variable quantification, sequential conjunction, and operation invocation, all expressed with readable keyword-based S-expressions instead of symbolic syntax. For example, a complex Narsese sentence like '<( light_on &/ <(* , { SELF } , switch ) --> ^ press >) =/> light_off >'—which encodes a temporal rule for turning off a light—becomes the clearer '( believe ( predict ( seq " light_on " ( call ^ press ( ext-set " SELF " ) " switch " ) ) " light_off " ) )' in DriftScript. The language's design focuses on readability, broad coverage of practical NAL constructs, static validation to catch errors early, and composability with raw Narsese input, making it easier for developers to create and debug agent programs.

The DriftScript compiler is implemented as a zero-dependency, four-stage pipeline in 1,941 lines of C99, consisting of tokenization, parsing, compilation, and emission stages. It processes source code into an abstract syntax tree with nodes for atoms and lists, enforcing arity constraints and validation rules for terms, copulas, and connectors. The compiler includes a comprehensive test suite of 106 cases across categories like tokenization, parsing, copulas, connectors, and error detection, all of which pass. Equivalence testing confirms that compiled DriftScript produces byte-identical engine output to hand-written Narsese, as demonstrated in examples like deduction chains and temporal rules, ensuring the transformation does not alter inference behavior. Structural readability metrics show that DriftScript reduces symbolic character usage by 36% and distinct symbol types by 60%, increasing the alphabetic ratio from 0.44 to 0.64, which supports qualitative claims of improved readability without changing the underlying reasoning logic.

This development has significant for building autonomous agents that interact with real-world environments. When integrated with the DriftNARS engine, DriftScript programs can connect to external systems through four structured callback types—event, answer, decision, and execution handlers—and an HTTP operation registry, enabling a sense-reason-act loop. For instance, in a worked example, an agent learns to associate seeing food with grabbing it to have food, using temporal learning and goal-directed action without explicit rules. Another example demonstrates multi-step planning with Python callbacks, where the agent picks up a key and unlocks a door based on state feedback. The HTTP-based agent example shows how operations can be registered and triggered via standard web protocols, making it feasible to deploy such agents in networked applications like smart home systems or robotics, where adaptability and real-time decision-making are crucial.

Despite its advancements, DriftScript has limitations: it is a syntactic transformation without formal semantics beyond compilation to Narsese, lacking type checking, semantic analysis, or optimization. The evaluation is structural and demonstrative, based on metrics and test cases, rather than comparative or user-study data, so its impact on actual developer productivity remains unquantified. Future work could include semantic validation for issues like eternal temporal , inline callbacks to bridge declarative and imperative programming, a language server for better tooling, and formal user studies to measure readability improvements. However, as a readable authoring surface, DriftScript does not modify NARS's core reasoning capabilities but makes them more accessible, potentially accelerating research and application in adaptive AI systems.