# Multi-Agent Workflow

6 related articles

Subscribe to updates

arXiv Oct 8, 2026
An author–reviewer loop matched a multi-agent workflow on MCFM Fortran-to-C++ translation at about one-third the cost per file
The work proposes collaborative human–AI agentic workflows in which domain experts write the specification and plan and agents write code under a deterministic orchestration pattern, with each stage ending in a numerical comparison against reference code and human approval; across fourteen experiments on translating MCFM from Fortran to C++, a simple author–reviewer loop with enforced limits settled a comparable number of files to a multi-agent workflow at about one-third the cost per file.
The work proposes collaborative human–AI agentic workflows in which domain experts write the specification and plan and agents write code under a deterministic orchestration pattern, with each stage ending in a numerical comparison against reference code and human approval; across fourteen experiments on translating MCFM from Fortran to C++, a simple author–reviewer loop with enforced limits settled a comparable number of files to a multi-agent workflow at about one-third the cost per file.
Akash Dhruv（First author）
Author–Reviewer Loop CodeScribe Multi-Agent Workflow Code Translation MCFM
arXiv Oct 2, 2026
InFlowOp uses one label-free cost to set both task decomposition and fault repair, gaining up to 11.97% over a single agent on Braid
The work proposes InFlowOp, which prices every decision in one label-free cost that weighs how well an agent's competence meets a subtask's demands against how much that agent takes to run, using it before execution to bidirectionally set decomposition granularity and agent assignment and during execution to locate and locally correct faults, and introduces Braid, a benchmark whose tasks require multi-agent coordination, reporting gains up to +11.97% over single-agent baselines across eight domains and six backbones, with +9.64% from in-flow optimization.
The work proposes InFlowOp, which prices every decision in one label-free cost that weighs how well an agent's competence meets a subtask's demands against how much that agent takes to run, using it before execution to bidirectionally set decomposition granularity and agent assignment and during execution to locate and locally correct faults, and introduces Braid, a benchmark whose tasks require multi-agent coordination, reporting gains up to +11.97% over single-agent baselines across eight domains and six backbones, with +9.64% from in-flow optimization.
Xuehang Guo（First author）
InFlowOp Braid Workflow Optimization Multi-Agent Workflow
arXiv Oct 2, 2026
FloWright treats the workflow as a training harness: co-evolving multiple roles lifts small open models by up to 7.41% across document, slide, chart, code, math, and finance tasks
The work proposes FloWright, which uses the workflow itself as a harness and a hierarchical, structure-aware reward to localize a workflow's single sparse outcome to individual roles, enabling one role to self-evolve and two or more roles to co-evolve, together with DataWright, which hardens existing single-agent datasets into workflow-level tasks; across document, slide, chart, code, math, and finance tasks, small open models trained with FloWright improve performance by up to 7.41%, with co-evolving (+5.03%) gaining more than optimizing one role alone (+2.83%).
The work proposes FloWright, which uses the workflow itself as a harness and a hierarchical, structure-aware reward to localize a workflow's single sparse outcome to individual roles, enabling one role to self-evolve and two or more roles to co-evolve, together with DataWright, which hardens existing single-agent datasets into workflow-level tasks; across document, slide, chart, code, math, and finance tasks, small open models trained with FloWright improve performance by up to 7.41%, with co-evolving (+5.03%) gaining more than optimizing one role alone (+2.83%).
University of Illinois Urbana-Champaign
Xuehang Guo（First author）
FloWright DataWright Multi-Agent Workflow
NVIDIA Technical Blog Sep 17, 2026
A multi-agent skill ported all 24 TileGym operators from cuTile Python/Triton-TileIR to cuTile Rust, averaging 99.5% of reference performance
The work built a multi-agent AI skill that translates GPU tile kernels written in cuTile Python and Triton-TileIR into cuTile Rust, and used it to port all 24 public TileGym operators (roughly 40 kernels, spanning element-wise operations through flash-attention decode, MLA, and MoE), measuring a geometric mean of 0.995 by CUPTI device time on NVIDIA DGX B200, i.e. 99.5% of cuTile Python performance on average, with every stage gated by machine-checkable verdicts and Tile IR diffs.
The work built a multi-agent AI skill that translates GPU tile kernels written in cuTile Python and Triton-TileIR into cuTile Rust, and used it to port all 24 public TileGym operators (roughly 40 kernels, spanning element-wise operations through flash-attention decode, MLA, and MoE), measuring a geometric mean of 0.995 by CUPTI device time on NVIDIA DGX B200, i.e. 99.5% of cuTile Python performance on average, with every stage gated by machine-checkable verdicts and Tile IR diffs.
NVIDIA
Multi-Agent Workflow cuTile Rust TileGym CUDA Tile IR
arXiv Sep 30, 2026
Meituan's LongCat team splits deep research into planning, parallel section research, and global-to-local editing via ResearchSpec, scoring 55.25, 51.35, and 79.83 on three public benchmarks
Meituan's LongCat team presents LongCat-DeepResearch, which shifts early research iteration from the full report to an executable ResearchSpec: multiple planning agents first search external sources and consolidate a research plan, researchers then investigate and draft citation-bearing sections in parallel independent contexts, and a Global Editor assigns cross-section ownership while Local Editors make targeted revisions, yielding 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics, and 76.04 on an in-house benchmark, second among four compared systems.
Meituan's LongCat team presents LongCat-DeepResearch, which shifts early research iteration from the full report to an executable ResearchSpec: multiple planning agents first search external sources and consolidate a research plan, researchers then investigate and draft citation-bearing sections in parallel independent contexts, and a Global Editor assigns cross-section ownership while Local Editors make targeted revisions, yielding 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics, and 76.04 on an in-house benchmark, second among four compared systems.
Meituan LongCat Team（First author）
LongCat-DeepResearch ResearchSpec Multi-Agent Workflow Deep Research
medRxiv Sep 21, 2026
MES: A Multi-Agent Evidence Synthesis System for Medical Decision-Making
This work presents MES, a multi-agent framework of six specialized agents (Supervisor, LiteratureMiner, RWD-Analyst, KG-Specialist, Statistician, SafeChecker) that integrates published literature and trial evidence, generates question-specific real-world evidence from real-world clinical data, and queries biomedical knowledge graphs while preserving source traceability; evaluated in two clinical use cases (an Alzheimer disease medication question with no matching literature and a septic shock beta-blocker question with inconsistent randomized evidence) and across 144 clinical queries spanning six evidence-based medicine categories, it produced structured reports that preserved source traceability, identified cross-source disagreement and communicated uncertainty, and its SafeChecker improv
This work presents MES, a multi-agent framework of six specialized agents (Supervisor, LiteratureMiner, RWD-Analyst, KG-Specialist, Statistician, SafeChecker) that integrates published literature and trial evidence, generates question-specific real-world evidence from real-world clinical data, and queries biomedical knowledge graphs while preserving source traceability; evaluated in two clinical use cases (an Alzheimer disease medication question with no matching literature and a septic shock beta-blocker question with inconsistent randomized evidence) and across 144 clinical queries spanning six evidence-based medicine categories, it produced structured reports that preserved source traceability, identified cross-source disagreement and communicated uncertainty, and its SafeChecker improv
Weill Cornell Medicine Cornell University
Haoyang Li（First author）Fei Wang（Corresponding author）
MES SafeChecker Multi-Agent Workflow Medical Evidence Synthesis Target Trial Emulation

You’re all caught up

An author–reviewer loop matched a multi-agent workflow on MCFM Fortran-to-C++ translation at about one-third the cost per file

InFlowOp uses one label-free cost to set both task decomposition and fault repair, gaining up to 11.97% over a single agent on Braid

FloWright treats the workflow as a training harness: co-evolving multiple roles lifts small open models by up to 7.41% across document, slide, chart, code, math, and finance tasks

A multi-agent skill ported all 24 TileGym operators from cuTile Python/Triton-TileIR to cuTile Rust, averaging 99.5% of reference performance

Meituan's LongCat team splits deep research into planning, parallel section research, and global-to-local editing via ResearchSpec, scoring 55.25, 51.35, and 79.83 on three public benchmarks

MES: A Multi-Agent Evidence Synthesis System for Medical Decision-Making