Agent-oriented software engineering (AOSE) is a software engineering paradigm that arose to apply best practice in the development of complex Multi-Agent Systems (MAS) by focusing on the use of agents, and organizations (communities) of agents as the main abstractions. The field of Software Product Lines (SPL) covers all the software development lifecycle necessary to develop a family of products where the derivation of concrete products is made systematically and rapidly.
Benchmarks Several benchmarks have been developed to evaluate the capabilities of AI coding agents and large language models in software engineering tasks. Here are some of the key benchmarks:
+ Agentic software engineering benchmarks Benchmark Description SWE-bench Assesses the ability of AI models to solve real-world software engineering issues sourced from GitHub repositories. The benchmark involves: Providing agents with a code repository and issue description Challenging them to generate a patch that resolves the described problem Evaluating the generated patch against unit tests ML-Agent-Bench Designed to evaluate AI agent performance on machine learning tasks τ-Bench τ-Bench is a benchmark developed by Sierra AI to evaluate AI agent performance and reliability in real-world settings. It focuses on: Testing agents on complex tasks with dynamic user and tool interactions Assessing the ability to follow domain-specific policies Measuring consistency and reliability at scale WebArena Evaluates AI agents in a simulated web environment. The benchmark tasks include: Navigating complex websites to complete user-driven tasks Extracting relevant information from the web Testing the adaptability of agents to diverse web-based challenges AgentBench A benchmark designed to assess the capabilities of AI agents in handling multi-agent coordination tasks. The key areas of evaluation include: Communication and cooperation between agents Task efficiency and resource management Adaptability in dynamic environments MMLU-Redux An enhanced version of the MMLU benchmark, focusing on evaluating AI models across a broad range of academic subjects and domains. It measures: Subject matter expertise across multiple disciplines Ability to handle complex problem-solving tasks Consistency in providing accurate answers across topics McEval A coding benchmark designed to test AI models' ability to solve coding challenges. The benchmark evaluates: Code correctness and efficiency Ability to handle diverse programming languages Performance across different coding paradigms and tasks CS-Bench A specialized benchmark for evaluating AI performance in computer science-related tasks. The key focus areas include: Algorithms and data structures Computational complexity and optimization Theoretical and applied computer science concepts WildBench Tests AI models in understanding and reasoning about real-world wild environments. It emphasizes: Handling noisy and unstructured data Adapting to unpredictable changes in the environment Performing well in multi-modal scenarios with real-world relevance Test of Time A benchmark that focuses on evaluating AI models' ability to reason about temporal sequences and events over time. It assesses: Understanding of temporal logic and sequence prediction Ability to make decisions based on time-dependent data Performance in tasks requiring long-term planning and foresight
Software engineering agent systems
There are several software engineering (SWE) agent systems in development. Here are some examples:
+ List of SWE Agent Systems SWE Agent System Backend LLM Salesforce Research DEIBASE-1 gpt4o Cosine Genie Fine-tuned OpenAI GPT CodeStory Aide gpt4o + Claude 3.5 Sonnet AbenteAI MentatBot gpt4o Salesforce Research DEIBASE-2 gpt4o Salesforce Research DEI-Open gpt4o Bytedance MarsCode gpt4o Alibaba Lingma gpt-4-1106-preview Factory Code Droid Anthropic + OpenAI AutoCodeRover gpt4o Amazon Q Developer (unknown) CodeR gpt-4-1106-preview MASAI (unknown) SIMA gpt4o Agentless gpt4o Moatless Tools Claude 3.5 Sonnet IBM Research Agent (unknown) Aider gpt4o + Claude 3 Opus OpenDevin + CodeAct gpt4o AgileCoder (various) ChatDev (unknown) MetaGPT gpt4o
References Michael Winikoff and Lin Padgham. Agent Oriented Software Engineering. Chapter 15 (pages 695-757) In G. Weiss (Ed.). Multiagent Systems. 2nd Edition. MIT Press. (a recent survey of the field) Site of the MaCMAS methodology which is applying MAS-PL. https://web.archive.org/web/20100922120209/http://james.eii.us.es/MaCMAS/index.php/Main_Page MAS Product Lines site: https://web.archive.org/web/20140518122645/http://mas-productlines.org/ Joaquin Peña, Michael G. Hinchey, and Antonio Ruiz-Cortés. Multiagent system product lines: Challenges and benefits. Communications of the ACM, December 2006, volume 49, issue number 12.
Joaquin Peña, Michael G. Hinchey, Antonio Ruiz-Cortés, and Pablo Trinidad. Building the Core Architecture of a NASA Multiagent System Product Line. In 7th International Workshop on Agent Oriented Software Engineering 2006, page to be published, Hakodate, Japan, May 2006. LNCS. https://doi.org/10.1007%2F978-3-540-70945-9_13 Joaquin Peña, Michael G. Hinchey, Manuel Resinas, Roy Sterritt, James L. Rash. Managing the Evolution of an Enterprise Architecture using a MAS-Product-Line Approach. 5th Int. Workshop on System/Software Architectures (IWSSA’06). Nevada, USA. 2006 Soe-Tsyr Yuan. MAS Building Environments with Product-Line-Architecture Awareness. Josh_Dehlinger and Robyn Lutz have several publications in this field. MAS-PL -- Current research. In THE FOURTH TECHNICAL FORUM (TF4) of AgentLink. December 2006.
External links Agent-Oriented Software Engineering: Reflections on Architectures, Methodologies, Languages, and Frameworks