2026 年 AI コーディングエージェント向けオープンソースベンチマーク 10 選

**

"What gets measured gets improved."

For years, coding benchmarks mostly measured one thing: could a model write a function that passed the unit tests? While that was useful, it doesn't reflect the reality of software engineering. Modern agentic coding benchmarks evaluate whether AI agents can work inside real repositories, edit existing code, run tests and other commands, debug failures, and maintain code quality as they complete longer and more complex tasks. They provide a much more realistic picture of how capable an AI coding agent really is. In this article, we'll look at the top 10 open-source agentic coding benchmarks** that you can use to evaluate your own model or simply keep track of where the field is headed. So, let's get started.

# 1. SWE-bench

**

SWE-bench** is still the benchmark most people think of first when evaluating agentic coding systems. It gives an AI agent a real GitHub issue and a snapshot of a real repository, then asks it to generate a patch that fixes the problem. To succeed, the agent has to understand the codebase, find the right files, make the required changes, and pass the tests. This benchmark includes 2,294 real software engineering tasks from 12 popular Python repositories. Variants like SWE-bench Lite and SWE-bench Verified have made it even more popular with researchers and model developers. While SWE-bench has become somewhat saturated and may no longer fully capture the challenges of modern coding agents, it remains the standard baseline. Almost every new coding agent is evaluated on SWE-bench, making it the benchmark that everyone tracks.

# 2. Terminal-Bench

**

Terminal-Bench** evaluates whether AI agents can operate in real terminal environments, not just produce patches or code snippets. These include compiling code, setting up servers, installing dependencies, running tests, debugging failures, and fixing security issues. This makes Terminal-Bench much closer to how developers actually work. It measures whether an agent can use shell commands, inspect files, read error messages, and iterate until the task is complete. The benchmark includes a task dataset and an execution harness that connects models to a sandboxed terminal, making it practical for reproducible evaluation. It ranks highly because terminal fluency is becoming a core requirement for serious coding agents. Newer editions like Terminal-Bench 2.1 include continuous validation, while Frontier-Bench represents the next evolution as a harder, domain-diverse successor.

# 3. SWE-Bench Pro

**

SWE-Bench Pro** is designed to test whether agents can solve more complex, long-horizon software engineering tasks. Like SWE-bench, it gives the agent a codebase and issue and asks for a patch, but the tasks are intended to be harder and more enterprise-like. The benchmark contains 1,865 problems from 41 actively maintained repositories, with a public split, a held-out split, and a commercial split. The tasks are designed to require larger patches, more context, and more realistic engineering work. It is especially useful if you want to know whether an agent can handle messy production-style problems rather than only small, well-scoped GitHub issues. However, a 2026 OpenAI audit has raised concerns about some dataset quality issues (~30%), including broken or overly strict test cases, highlighting the need for careful benchmark validation.

# 4. Senior SWE-Bench

**

Senior SWE-Bench** focuses on a gap that most coding-agent benchmarks miss: real senior engineers are not only judged on whether code runs, but also on maintainability, design judgment, and how well their code aligns with the surrounding codebase. The benchmark evaluates agents on senior-level work such as feature implementation, bug investigation, performance work, and codebase-convention alignment. Its initial release contains 100 tasks across 12 open-source repositories, with public and private splits, and every task goes through multiple review layers. This is important because future coding agents will need to produce code that teams actually want to merge, not just patches that pass a narrow verifier.

# 5. Agents' Last Exam

**

Agents' Last Exam** (ALE) evaluates the broader class of long-horizon agents that coding systems are becoming. ALE measures AI agents on economically valuable professional workflows with verifiable outcomes, covering 55 subdomains across 13 industry clusters. It was recently featured in GPT-5.6 Sol's release, where it set a new high of 53.6. The public GitHub project and metadata release make it useful for researchers studying whether agents can complete realistic work, not just solve benchmark-shaped tasks. For coding-agent teams, its value is that it pushes evaluation toward full professional workflows, where software work is often mixed with data handling, terminal use, research, file manipulation, and verification.

# 6. DeepSWE

**

DeepSWE** is a newer benchmark for frontier coding agents built around original long-horizon engineering tasks from active open-source repositories. It includes 113 tasks across TypeScript, Go, Python, JavaScript, and Rust, with isolated environments and program-based verifiers. Its major advantage is that the tasks are written from scratch rather than simply mined from public merged fixes, reducing the chance that models are reconstructing memorized solutions. It also grades observable behavior rather than requiring one exact reference patch. This makes DeepSWE one of the most promising benchmarks for measuring genuine software engineering ability rather than benchmark recall.

# 7. SlopCodeBench

**

SlopCodeBench** measures what happens when coding agents repeatedly extend their own previous solutions as requirements evolve. That is important because many real agentic workflows are iterative: an agent writes code, then modifies it, then extends it again. Single-shot benchmarks often miss the quality degradation that can happen over multiple rounds. SlopCodeBench is designed to expose path dependence, non-convergence, and structural instability under iterative specification refinement. The benchmark is language-agnostic and includes 20 problems with 93 checkpoints, tracking not only correctness but also verbosity and structural erosion. This is valuable because agentic coding is increasingly used in iterative development, where early architectural decisions compound over time.

# 8. Multi-SWE-bench

**

Multi-SWE-bench** extends the SWE-bench idea beyond Python. It focuses on multilingual issue resolution across Java, TypeScript, JavaScript, Go, Rust, C, and C++, with 1,632 high-quality instances annotated from a larger candidate pool. This matters because real coding agents cannot be evaluated only on Python repositories. A useful agent needs to understand package managers, build systems, language idioms, test frameworks, and repository layouts across many ecosystems. Multi-SWE-bench also releases data, code, and container environments, making it useful for open research and reproducible evaluation.

# 9. ProgramBench

**

ProgramBench** asks whether agents can rebuild complete programs from scratch. Given only a compiled binary and documentation, the agent must architect and implement a codebase that reproduces the program's behavior. That is a very different skill from fixing a GitHub issue or editing one file. It tests architecture, abstraction, implementation planning, and behavioral matching. The benchmark is especially useful because it evaluates whole-program development rather than narrow patch generation. The benchmark includes 200 tasks and uses agent-driven fuzzing to generate more than 248,000 behavioral tests. It is relatively new but provides a valuable measure of an agent's ability to build coherent software systems end to end.

# 10. Aider Polyglot

**

Aider Polyglot** evaluates how well models can follow instructions and edit code across C++, Go, Java, JavaScript, Python, and Rust. It uses 225 challenging Exercism coding exercises and tests not only first-attempt coding but also the ability to respond to failing unit tests and repair the solution. It is smaller and more coding-exercise-like than repository-level benchmarks, but it is useful because it is easy to run, language-diverse, and practical for comparing code-editing ability across models.

# Wrapping Up

AI coding agents are improving quickly, but measuring their progress is becoming just as important as building the agents themselves. No single benchmark tells the complete story. SWE-bench tests issue resolution, Terminal-Bench evaluates real-world tool usage, ProgramBench explores full program generation, and newer benchmarks focus on areas like long-horizon planning, maintainability, and iterative development. As AI agents become more capable, benchmarks will need to continue evolving alongside them.

Kanwal Mehreen is a machine learning engineer and a technical writer with a profound passion for data science and the intersection of AI with medicine. She co-authored the ebook "Maximizing Productivity with ChatGPT". As a Google Generation Scholar 2022 for APAC, she champions diversity and academic excellence. She's also recognized as a Teradata Diversity in Tech Scholar, Mitacs Globalink Research Scholar, and Harvard WeCode Scholar. Kanwal is an ardent advocate for change, having founded FEMCodes to empower women in STEM fields.