cs.AIFeb 26, 2026
SaveMiroFlow: Towards High-Performance and Robust Open-Source Agent Framework for General Deep Research Tasks
Organizations: Tsinghua University · MiroMind AI · National University of Singapore · Nanjing University
Abstract
Despite the remarkable progress of large language models (LLMs), the capabilities of standalone LLMs have begun to plateau when tackling real-world, complex tasks that require interaction with external tools and dynamic environments. Although recent agent frameworks aim to enhance model autonomy through tool integration and external interaction, they still suffer from naive workflows, unstable performance, limited support across diverse benchmarks and tasks, and heavy reliance on costly commercial APIs. In this work, we propose a high-performance and robust open-source agent framework, termed MiroFlow, which incorporates an agent graph for flexible orchestration, an optional deep reasoning mode to enhance performance, and a robust workflow execution to ensure stable and reproducible performance. Extensive experiments demonstrate that MiroFlow consistently achieves state-of-the-art performance across multiple agent benchmarks, including GAIA, BrowseComp-EN/ZH, HLE, xBench-DeepSearch, and notably FutureX. We hope it could serve as an easily accessible, reproducible, and comparable baseline for the deep research community.
Figures & tables
Figure 1: Overall performance of MiroFlow on representative deep research benchmarks. MiroFlow, a high-performance and robust open-source agent framework, achieves reproducible state-of-the-art results across all benchmarks, consistently outperforming existing open-source and commercial agent systems. All MiroFlow results are obtained with a single unified configuration without any task-specific tuning, demonstrating strong generality and adaptability across heterogeneous deep research tasks.
Figure 2: Overview of the three-tier hierarchical MiroFlow framework architecture. Foundation tier provides reusable core components, including LLM backends, MCP-based tool sets, and generic input–output processors, which supply the basic capabilities required by all agents. Agent tier defines a set of agent nodes constructed by combining foundation-tier components with specific prompts. Each node is provided with a list of agents it can call. These nodes communicate through structured messages and can be flexibly instantiated or extended. Control tier assembles multiple agent nodes into an agent graph and orchestrates the end-to-end workflow: user queries enter the graph and are processed through coordinated agent interactions and tool calls, while the controller maintains task logs and checkpoints for reproducibility, supports a heavy-reasoning mode to improve accuracy, and incorporates workflow-level robustness enhancements to ensure smooth and predictable execution.
Figure 3: Illustration of Heavy-Reasoning Mode.
| Agent Framework | LLM Base Model | GAIA Val | GAIA Test | BrowseComp | HLE | HLE Text | xBench -DS | ||
| AVG | Text | EN | ZH | ||||||
| Closed-source frameworks | |||||||||
| Manus [ 32 ] | - | 73.3 | - | - | - | - | - | - | - |
| MiniMax-M2 [ 2 ] | - | - | 75.7 | - | 44.0 | 48.5 | 31.8 | - | 72.0 |
| OpenAI-DR [ 42 ] | - | 67.4 | - | - | 51.5 | 42.9 | - | 26.6 | - |
| Alita [ 48 ] | Claude–4, GPT–4o | 75.15 | - | 75.4 | - | - | - | - | - |
Table 1: Comparison of agent frameworks. As noted in previous studies [ 78 ] , some open-source frameworks are difficult to reproduce, and some reported performance remains ambiguous. Scores marked with an asterisk (*) indicate that the final answer is obtained by integrating the results of multiple independent agents. All other MiroFlow results are obtained without using heavy-reasoning mode. MiroThinker does not possess multimodal capabilities; therefore, evaluation results are not available (N/A) on GAIA-Val, GAIA-Test, and HLE benchmarks. “DS” in this table and throughout the paper refers to DeepSearch, and “DR” refers to DeepResearch.
| Agent Framework | Base LLM | FutureX |
| MiroFlow (ours) | GPT-5 | 42.5 |
| GPT-5-Thinking-Heavy [ 40 ] | GPT-5 | 25.6 |
| ChatGPT-Agent [ 39 ] | - | 21.9 |
| Qwen3-235B [ 69 ] | Qwen3-235B | 19.0 |
| EventDeepResearch-V2 [ 52 ] | Grok-4 | 18.7 |
| Deepseek-R1 [ 15 ] | Deepseek-R1 | 18.4 |
Table 2: Performance of Mainstream Agents on FutureX. These scores are taken from the official FutureX leaderboard [ 7 ] (November, Week 2), where MiroFlow has held the top position since September 2025. All models are evaluated with search tools.
| Setting | GAIA-Val | Std. Dev. (%) |
| MiroFlow (default) | 71.9 | 1.21 |
| w/o Message Normalization | 68.5 | 2.43 |
| w/o Retry Mechanism | 69.0 | 1.70 |
Table 3: Ablation results on Gaia benchmark under different settings. Message normalization and the retry mechanism enhance the agent’s stability and average performance.
| Setting | LLM Models | GAIA-Val |
| None (default) | GPT-5 | 71.9 |
| Ensemble | GPT-5 | 72.3 |
| GPT-5 + Claude-3.7 | 71.6 | |
| GPT-5 (diff. prompts) | 72.5 | |
| GPT-5 | 74.6 | |
| GPT-5 + Claude-3.7 | 73.0 |
Table 4: Heavy-Reasoning Mode ablations on GAIA-Val . Heavy-reasoning mode scales computational resources and leads to improved performance.
| Setting | GAIA-Val | BC-200 | HLE-200 |
| Single-Agent | 74.8 | 63.9 | 40.6 |
| Multi-Agent | 71.9 | 68.3 | 42.0 |
Table 5: Single- vs. Multi-Agent ablation within MiroFlow. Multi-agent settings perform better on most benchmarks but not on GAIA.
Figure 4: Accuracy vs. Max Turns on GAIA validation. Accuracy improves as max turn increases and then saturates. More difficult problems require more turns. Multi-agent settings saturate earlier but perform slightly worse than the single-agent setting.
| Setting | GAIA-Val | GAIA-Val-Text |
| default tool set | 71.9 | 79.9 |
| open-source tool set | 70.3 | 79.0 |
Table 6: Comparison of MiroFlow performance under different tool set configurations. Although open-source tools may not match the quality of commercial counterparts, they still achieve strong performance when integrated with MiroFlow.
| Setting | Avg. (%) | Std. Dev. (%) |
| No processing | 59.60 | 3.34 |
| Input only | 62.63 | 0.70 |
| Output only | 73.94 | 2.18 |
| Input + Output (default) | 71.9 | 1.21 |
Table 7: Pass@1 results under different I/O processing settings on GAIA-Validation benchmarks. I/O processing helps the agent focus on the intended reasoning path and reduces instruction-following errors.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Open-sourced Agent Framework | LLM Base Model | Flexibility | Stability | GAIA Val Avg |
| smolagent [ 51 ] | Openai o1 | Single-Agent | Good | 49.7 |
| OWL [ 21 ] | Claude–3.7–Sonnet | Multi-Agent | Poor | 69.7 |
| AWorld [ 72 ] | Claude–3.7–Sonnet | Multi-Agent | Fair | - |
| Tongyi–D.R. [ 60 ] | Tongyi–D.R.–30B–A3B | Single-Agent | Fair | 70.9 |
| AgentOrchestra [ 76 ] | Claude–3.7–Sonnet, GPT–4.1 OpenAI–Computer–Use, etc. | Multi-Agent | Poor | 82.4 |
| JoyAgent [ 30 ] | Claude-4, o4-mini | Multi-Agent | Fair | 75.15 |
Table 8: Comparison of agent frameworks. Flexibility : “Single-Agent” and “Multi-Agent” indicates single-agent or multi-agent frameworks, some of them are designed for specific benchmarks; “Graph” denotes general frameworks supporting dynamic and adaptive agent graph workflows across diverse tasks. Stability : “Poor” indicates frameworks that are difficult to reproduce; “Fair” denotes reproducible performance under well-maintained environments; “Good” represents frameworks with special designs enhancing robustness. Scores marked with an asterisk (*) indicate that the final answer is obtained by integrating the results of multiple independent agents. All MiroFlow results are obtained without using heavy-reasoning mode. “D.R.” in this table refers to DeepResearch.
| Framework | GAIA Val (pass@1) | |||
| L1 | L2 | L3 | Avg | |
| OWL (Reported) | 84.9 | 68.6 | 42.3 | 69.7 |
| OWL (Reproduced) | 78.8 | 47.7 | 26.9 | 53.9 |
| Performance Difference | -6.1 | -20.9 | -15.4 | -15.8 |
| AgentOrchestra (Reported) | 92.5 | 83.7 | 57.7 | 82.4 |
| AgentOrchestra (Reproduced) | 62.3 | 54.7 | 38.5 | 54.6 |
Table 9: Reproducing results for OWL and AgentOrchestra. L1, L2, and L3 represent three difficulty levels of the GAIA Validation set, ordered by increasing difficulty.
Figure 12: Mistake propagation in multi-agent. Multi-agent structure has multiple different and separate context. Decomposing the sequential tasks has the risk of information loss or mistake propagation across agents, whereas a single-agent model has a more stable global context. The red bounding box highlights where the agent made an error (caused by using a low-resolution image that produced a biased result). Because a single agent maintains more comprehensive contextual awareness, it can further correct and verify this error in subsequent tasks. In contrast, the downstream agent in a multi-agent setting may lack access to such fine-grained details, leading it to accept the incorrect conclusion.
Figure 13: Single-Agent Baseline. The system includes only one Main Agent, which must repeatedly perform reasoning steps, trigger tools, and call itself to complete a complex task. Due to the lack of role decomposition or specialized subagents, this architecture exhibits minimal flexibility and struggles with tasks that require modular expertise or parallel processing.
Figure 14: Multi-Agent Baseline. The system consists of a Main Agent that decomposes a complex task into several subtasks and delegates them to multiple specialized Sub-Agents. Each Sub-Agent focuses on a specific type of reasoning or operation, enabling modularity, improved flexibility, and higher execution efficiency. Compared with the single-agent setup, this architecture better supports specialization, parallel processing, and robust task handling. However, it is still not flexible enough.
Figure 15: Agent Graph in short video generation. The goal of the system is to transform a user-provided script into a complete short-video asset package. The Main Agent plans the workflow and dispatches subtasks to specialized Sub-Agents. Sub-Agent 1 optimizes and refines the input text, Sub-Agent 2 generates a structured storyboard, Sub-Agent 3 retrieves or creates visual materials, Sub-Agent 4 produces dubbing audio, and Sub-Agent 5 generates subtitles and handles typesetting. The outputs from all agents are finally aggregated to produce the final video content.
Figure 16: Agent Graph in complex travel planning and reservation drafting. The goal of the system is to convert user constraints and travel preferences into a detailed multi-city itinerary and a reservation-ready plan. The Main Agent decomposes the task and coordinates specialized Sub-Agents. Sub-Agent 1 analyzes constraints such as time, budget, and travel rules; Sub-Agent 2 generates feasible candidate routes and transportation options; Sub-Agent 3 evaluates routes with scoring criteria; Sub-Agent 4 expands the selected plan into a detailed day-by-day itinerary; and Sub-Agent 5 drafts preliminary reservation information including flights, hotels, and activities. These outputs are merged to form the final travel plan delivered to the user.
Figure 17: Agent Graph in job application material generation and delivery strategy. The goal of the system is to transform a user’s background and target job preferences into a complete set of application materials and a personalized job-hunting strategy. The Main Agent plans the overall workflow and delegates subtasks to specialized Sub-Agents. Sub-Agent 1 analyzes the user’s background and extracts key experience, Sub-Agent 2 interprets job requirements and target role expectations, Sub-Agent 3 generates tailored resumes, Sub-Agent 4 produces customized cover letters, Sub-Agent 5 retrieves suitable job openings and recommendations, and Sub-Agent 6 formulates an optimized application strategy. All outputs are finally aggregated to produce the final job application package.
Figure 18: Agent performance with respect to LLM context length. Performance of MiroFlow on the GAIA validation set using GPT-5 in a single-agent configuration under different maximum-context settings (8k–400k tokens). Accuracy improves substantially when increasing context from 8k to the 24k–48k range, after which gains begin to saturate. Harder tasks (L3) exhibit the strongest dependence on longer contexts.
More in this topic
Autonomous Research Agents
Autonomous research agents use language models and tools to search literature, design and run experiments, analyze results, and support scientific discovery.
See all 318 papers