PaperlayerView exact revision on arXiv

Source-linked AI summary

GEAR: Feature-Based Configuration and Code Generation for Agentic Systems

Brell Sanwouo, Nada Zine, Paul Temple, Clément Quinton, Romain Rouvoy

arXiv:2610.35592v1cs.MAcs.SE

TL;DR

GEAR addresses the challenge of configuring MAS across frameworks with different abstractions and APIs by providing a shared, feature-based specification for validation and project generation. Its preliminary mapping assessment identified translation strategies for 271 of 320 evaluated feature–framework pairs (84.69%).

  • Problem

    MAS frameworks expose similar capabilities through different abstractions and APIs, requiring developers and researchers to navigate framework-specific concepts and configuration choices.

    Source: §I · p. 1

  • Method

    GEAR uses a common feature-based specification for agents, modules, workflows, and configuration choices, validates it, and generates framework-specific projects through connectors.

    Source: §I · p. 1, A. GEAR Core Features and Workflow · p. 2, B. GEAR Architecture and Implementation · p. 3

  • Results

    84.69% of evaluated feature–framework pairs had identified translation strategies: 271 of 320 pairs across 10 frameworks.

    Source: §IV · p. 4

  • Takeaways & Limitations

    Researchers can systematically vary MAS designs and reproduce experimental setups across frameworks without manually reimplementing each system.

    Source: §I · p. 1

    Multimodal inputs and streaming are out of scope, and unresolved properties are omitted, so generated projects may differ from the requested specification.

    Source: §V · p. 4

Abstract

from arXiv · show

Engineering multi-agent systems (MAS) requires choosing agents, orchestration strategies, communication mechanisms, reasoning approaches, and memory configurations. These choices depend on framework-specific capabilities and abstractions, requiring developers to learn different APIs when exploring alternative implementations. We present GEAR, an open-source tool for configuring MAS and generating framework-specific implementations from a common specification. GEAR uses a feature-based representation to make configuration options and their constraints explicit, enabling users to validate specifications before code generation. Through connectors for ten agentic frameworks, users generate projects from the same specification and can then inspect reports describing framework-specific adaptations. A preliminary assessment of mapping feasibility identifies translation strategies for 271 of 320 evaluated feature-framework pairs (84.69%). GEAR supports developers prototyping agentic systems and researchers exploring MAS configurations across frameworks. GEAR is available as open-source software at https://github.com/brellsanwouo/gear-framework and a video demo is available at https://youtu.be/iwNi-oC2tCw.

I. INTRODUCTION

GEAR addresses the difficulty of configuring multi-agent systems across frameworks whose similar capabilities are exposed through different abstractions and APIs. It provides a common specification that can be validated and used to generate framework-specific implementations.

I. INTRODUCTION

  • Different framework abstractions and APIs make MAS design time-consuming, error-prone, and difficult to reason about.Developers and researchers must also navigate a large configuration space and learn framework-specific concepts and implementation choices.Source: §I · p. 1
  • Feature Models represent configuration choices and constraints explicitly and support automated validation, but the authors knew of no tool applying them to MAS variability.This builds on variability modeling used in configurable software systems.Source: §I · p. 1
  • GEAR lets users specify agents, modules, workflows, and choices once, validate the specification, and generate implementations for selected frameworks.Its connector-based architecture separates the shared specification from implementation logic, allowing integration of emerging frameworks and capabilities.Source: §I · p. 1
  • 84.69% of evaluated feature–framework pairs had identified translation strategies across GEAR’s 10 supported frameworks.The assessment identified strategies for 271 of 320 pairs.Source: §I · p. 1
  • GEAR targets developers prototyping LLM-based MASs and researchers studying agentic designs, including reuse, comparison, and reproducibility across frameworks.Researchers can vary MAS designs and reproduce experimental setups without manually reimplementing each system.Source: §I · p. 1

II. RELATED WORK

Prior MAS frameworks commonly embed design decisions in code, while declarative specifications offer limited support for modeling variability across alternative designs. GEAR addresses variability at the MAS level across heterogeneous frameworks.

Specification of MAS

  • MAS frameworks such as CAMEL-AI, AutoGen, MetaGPT, CrewAI, and Google ADK mainly use code-level definitions that embed design decisions in implementation artifacts.These approaches offer limited support for reuse or systematic exploration of alternative configurations.Source: §II · p. 1
  • Open Agent Specification, ADL, AgentSPEX [14], and H¨arer separate design from execution, but mainly describe individual configurations or execution flows.The passage characterizes their support for explicit variability across alternative MAS designs as limited.Source: §II · p. 1

Variability Modeling

  • Variability modeling has been applied to software systems, machine learning systems, LLM configurations [17], and industrial AI agents.Yoshimura et al. study adapting a domain-specific agent across operational contexts; GEAR instead models variability at the MAS level across frameworks.Source: §II · pp. 1–2

A. GEAR Core Features and Workflow

GEAR Studio guides users from a framework-independent MAS specification through validation and framework selection to generated projects. The workflow example uses two connected agents and reuses their specification to generate projects for different frameworks.

Configure

  • Users define agents, reusable orchestration modules, and workflows that connect components and specify execution flow.The log-analysis example groups LogParser and LogAnalyzer into a DiagnosisLoop module.Source: A. GEAR Core Features and Workflow · p. 2

Validate

  • GEAR checks agents, modules, and workflows for errors such as undefined references or incompatible configuration choices before framework selection.In the example, selecting both parallel and loop execution for DiagnosisLoop triggers a conflict because the strategies are mutually exclusive.Source: A. GEAR Core Features and Workflow · p. 2

Select Target Frameworks

Generate and Inspect and Execute

  • Each generated project contains source code, a configuration file, and an execution entry point.Users can inspect projects and execute them through the command line or through GEAR when execution is available.Source: A. GEAR Core Features and Workflow · p. 2

B. GEAR Architecture and Implementation

GEAR separates a common feature-based MAS specification from the processing components that produce framework-specific implementations. Its model encodes configuration options and constraints, while validation and connectors prepare each specification for deterministic assembly.

B. GEAR Architecture and Implementation

  • The specification layer uses feature modeling to represent configurable properties of agents, modules, and workflows, while the processing layer transforms specifications into framework-specific implementations.Feature Models organize properties hierarchically and encode mandatory, optional, or alternative choices and constraints.Source: B. GEAR Architecture and Implementation · pp. 2–3, Figure 2 · p. 3
  • GEAR’s feature model has 53 features and encodes approximately 2.4×10^8 valid configurations.The model covers agent identity, model and task configuration, tools, memory, reasoning, execution controls, and orchestration; 32 features were identified through literature review and analysis of ten frameworks, [5], [6].Source: B. GEAR Architecture and Implementation · p. 3, Figure 3 · p. 3
  • Validation checks specification structure, feature-model constraints, and workflow consistency, then normalizes valid configurations into a common intermediate representation.Framework connectors apply explicit feature mappings and provide manifests, templates, and assembly logic; the Assembly engine deterministically generates executable artifacts, and conversion reports record mapped, adapted, or unresolved features.Source: B. GEAR Architecture and Implementation · p. 3

IV. PRELIMINARY MAPPING ASSESSMENT

The assessment examined whether GEAR features could be mapped to ten frameworks, identifying strategies for most evaluated feature–framework pairs. A user study is ongoing to evaluate usability, effort, and participant difficulties.

IV. PRELIMINARY MAPPING ASSESSMENT

  • Mappings counted as identified when implementation was feasible through a native mechanism, adaptation, or synthesized code; otherwise they were unresolved.Unresolved cases included features not exposed by a framework or implicit in its internal mechanisms.Source: §IV · p. 4
  • 271 of 320 feature–framework pairs (84.69%) had identified mappings, while 49 (15.31%) remained unresolved.CrewAI had mappings for 31 features (96.88%), and Google ADK for 30 (93.75%); Microsoft Agent Framework and PydanticAI had mappings for 24 (75.00%) and 23 (71.88%), respectively.Source: §IV · p. 4
  • An ongoing user study evaluates usability, effort, and the difficulties participants encounter while specifying and generating MASs with GEAR.Source: §IV · p. 4

V. LIMITATIONS & FUTURE WORK

GEAR’s current scope centers on shared framework abstractions, and adding frameworks requires manually defined connectors and mappings. The authors call for further study of behavioral differences, broader feature support, and evaluation on larger real-world applications.

V. LIMITATIONS & FUTURE WORK

  • Advanced capabilities such as multimodal inputs and streaming are out of scope because they are highly framework-specific and were excluded to preserve a common abstraction level.Source: §V · p. 4
  • Adding a target framework requires manually defining connectors and mappings, initially demanding engineering effort and knowledge of that framework.The mappings are reusable once implemented; partially automated connector generation or mapping inference is proposed as future work.Source: §V · p. 4
  • Unresolved properties are omitted, so generated projects may differ from the requested specification; the authors call for more systematic study of these behavioral differences.Source: §V · p. 4
  • The authors identify evaluation on larger and more representative real-world multi-agent applications as future work.Source: §V · p. 4

VI. CONCLUSION

GEAR separates MAS specifications from framework-specific implementations, supporting reuse across heterogeneous frameworks through connectors. Its feature-based model makes configuration choices explicit, while its connector-based design supports adding frameworks.

VI. CONCLUSION

  • GEAR is an open-source tool for configuring and generating MASs across heterogeneous agentic frameworks.Source: §VI · p. 4
  • Separating MAS specifications from framework-specific implementations lets users reuse one specification across frameworks through connectors.Source: §VI · p. 4
  • GEAR’s feature-based model makes configuration choices explicit, and its connector-based design makes support for new frameworks easier.Source: §VI · p. 4
Loading 2610.35592v1…