cs.SEApr 8, 2026

ReCodeAgent: A Multi-agent Workflow for Language-Agnostic Translation and Validation of Large-Scale Repositories

Authors: Ali Reza Ibrahimzada, Brandon Paulsen, Daniel Kroening, Reyhaneh Jabbarvand

Organizations: University of Illinois Urbana-Champaign Urbana, IL, USA · Amazon Arlington, VA, USA · Amazon Seattle, WA, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Most repository-level code translation and validation techniques have been evaluated on a single source-target programming language (PL) pair, owing to the complex engineering effort required to adapt new PL pairs. Programming agents can enable PL-agnosticism in repository-level code translation and validation: they can synthesize code across many PLs and autonomously use existing tools specific to each PL's analysis. However, state-of-the-art has yet to offer a fully autonomous agentic approach for repository-level code translation and validation of large-scale programs. This paper proposes ReCodeAgent, an autonomous multi-agent approach for language-agnostic repository-level code translation and validation. Users only need to provide the project in the source PL and specify the target PL for ReCodeAgent to automatically translate and validate the entire repository. ReCodeAgent is the first technique to achieve high translation success rates across many PLs. We compare the effectiveness of ReCodeAgent with four alternative neuro-symbolic and agentic approaches to translate 118 real-world projects, with 1,975 LoC and 43 translation units for each project, on average. The projects cover 6 PLs and 4 PL pairs. Our results demonstrate that ReCodeAgent consistently outperforms prior techniques on translation correctness, improving test pass rate by 60.8% on ground-truth tests, with an average cost of $15.3. We also perform process-centric analysis of ReCodeAgent trajectories to confirm its procedural efficiency. Finally, we investigate how the design choices (a multi-agent vs. single-agent architecture) influence ReCodeAgent performance: on average, the test pass rate drops by 40.4%, and trajectories become 28% longer and persistently inefficient.

Figures & tables

Figure 1. Overview of ReCodeAgent .
Figure 3. Hover tool from the Python LSP server.
Figure 4. Project analysis (PA) tools output.
Figure 5. Documents generated by Analyzer Agent in ReCodeAgent .
Figure 6. Documents generated by Planning Agent in ReCodeAgent .
Tool (PL Pair)ProjectLoCCS (%)# Validated Developer TestsValidated Developer TestsReCodeAgent Translated Developer TestsReCodeAgent Generated TestsFunction Validation
TETPTFTETPTFTETPTFC (%)C+ (%)TotalSuccessFail
Oxidizer (Go →\rightarrow Rust)checkdigit428⟨\langle 100, 100 ⟩\rangle36⟨\langle 36, 36 ⟩\rangle⟨\langle 33, 36 ⟩\rangle⟨\langle 3, 0 ⟩\rangle363607171079.794.729⟨\langle 21, 29 ⟩\rangle⟨\langle 8, 0 ⟩\rangle
go-edlib639⟨\langle 100, 100 ⟩\rangle36⟨\langle 36, 36 ⟩\rangle⟨\langle 19, 36 ⟩\rangle⟨\langle 17, 0 ⟩\rangle3636033094.794.924⟨\langle 18, 24 ⟩\rangle⟨\langle 6, 0 ⟩\rangle
histogram314⟨\langle 100, 100 ⟩\rangle2⟨\langle 2, 2 ⟩\rangle⟨\langle 2, 2 ⟩\rangle⟨\langle 0, 0 ⟩\rangle2206666038.090.519⟨\langle 12, 19 ⟩\rangle⟨\langle 7, 0 ⟩\rangle
nameparts413⟨\langle 100, 100 ⟩\rangle26⟨\langle 26, 26 ⟩\rangle⟨\langle 23, 26 ⟩\rangle⟨\langle 3, 0 ⟩\rangle262602222096.896.815⟨\langle 9, 14 ⟩\rangle⟨\langle 6, 1 ⟩\rangle
stats1241⟨\langle 100, 100 ⟩\rangle121⟨\langle 121, 121 ⟩\rangle⟨\langle 71, 121 ⟩\rangle⟨\langle 50, 0 ⟩\rangle1211210320320043.379.652⟨\langle 38, 52 ⟩\rangle⟨\langle 14, 0 ⟩\rangle
Table 1. Effectiveness of ReCodeAgent in repository-level code translation and validation in terms of test and function validation. LoC: Lines of Code, CS: Compilation Success, TE: # Tests Executed, TP: # Tests Passing, TF: # Tests Failing, CRUST-{ α,β,σ,γ\alpha,\beta,\sigma,\gamma } : α\alpha : both compile, β\beta : only ReCodeAgent compile, σ\sigma : only SWE-agent compile, γ\gamma : both do not compile, C: Test coverage, C+: Test coverage with generated tests . Tuple entries indicate ⟨\langle Tool, ReCodeAgent ⟩\rangle .
ToolProject# Tests# Tests Translated / Not Translated# Tests w/ Matching / Non-Matching # Assertions# Total / Matching assertEqual OutputAssertion Type Match (%)Avg. Cosine SimilarityAvg. LoCAvg. # Method Invocations
Assert EqualAssert TrueAssert FalseOther
Oxidizer (Go →\rightarrow Rust)checkdigit3636/036/045/45100---0.94⟨\langle 24.42, 24.58 ⟩\rangle⟨\langle 7.14, 5.33 ⟩\rangle
go-edlib3636/036/045/4584.44---0.92⟨\langle 30.78, 51 ⟩\rangle⟨\langle 7.61, 9.50 ⟩\rangle
histogram22/02/011/11100---0.96⟨\langle 23.50, 16.50 ⟩\rangle⟨\langle 24, 16.50 ⟩\rangle
nameparts2626/026/051/51100---0.94⟨\langle 11.96, 6.31 ⟩\rangle⟨\langle 6.15, 3.81 ⟩\rangle
stats121121/0121/0150/15091.15---0.85⟨\langle 16.82, 9.44 ⟩\rangle⟨\langle 8.12, 6 ⟩\rangle
Table 2. Comparison between ReCodeAgent translated tests and original source PL tests. Tuple entries indicate ⟨\langle Source Test, Translated Test ⟩\rangle . LoC: Lines of Code.
Figure 7. Impact of different agents on translation effectiveness and agent trajectories. RA: ReCodeAgent , NoA: No Analyzer, NoP: No Planning, NoV: No Validator, BA- α\alpha : Base Agent with Prompt Condensation, BA- β\beta : Base Agent with Prompt Concatenation, NC: Node Count, TEC: Temporal Edge Count, SEC: Structural Edge Count, LC: Loop Count, ALL: Average Loop Length.
Figure 8. Cost and Tool Usage Analysis of ReCodeAgent .
ToolInput TokensOutput TokensTime (s)Cost ($)
Oxidizer131085.044855.2940.41.1
AlphaTrans2302094.0296475.237503.411.4
Skel45763.510733.6450.80.3
Crust1060519.411899.7455.14.0
ReCodeAgent446829.7257879.03433.215.3
Table 3. Cost comparison of existing techniques and ReCodeAgent averaged by number of projects translated.

Explore similar work

CardsList

  1. Zero2Repo: Can Coding Agents Build Repositories from Scratch?

    Sep 29, 2026Pei Yang, Tianyu Shi, Yuhang Yao +23Coding AgentsAI Agent Benchmarks

    Gradient Data · McGill University · Carnegie Mellon University +6

  2. Large-scale Repository Engineering via Agent-Native Reusable Code Primitives

    Oct 6, 2026Haibo Jin, Peng Kuang, Xucheng Yu +3Repository-Level Code GenerationSoftware Engineering

    University of Illinois Urbana-Champaign, USA

  3. Entropy-based Code Adversarial Translation for Real-world Repository Migration

    Aug 10, 2026Yushun Tang, Yisen Cao, Zhicheng Chen +4Code TranslationSoftware Engineering Agents

    Huawei Technologies Co., Ltd. · The Chinese University of Hong Kong