AI ranking
Coding agents
Development-oriented agents (Claude Code, Cursor Agent, Windsurf/Cascade, Aider). Assessed on their ability to modify, test and repair code.
Visual top 10
- 01 Live SWE-agent + claude 4.5 opus medium (20251101) 79.21
- 02 Sonar foundation agent + claude 4.5 opus 79.20
- 03 Trae + doubao Seed Code 78.81
- 04 OpenHands + claude opus 4.5 77.61
- 05 Atlassian rovo dev (2025 09 02) 76.81
- 06 Epam ai/run developer agent v20250719 + claude 4 sonnet 76.80
- 07 Acoder 76.41
- 08 Harness ai 74.81
- 09 Joycode + claude 4 sonnet + gpt 4.1 74.61
- 10 Lingxi V1.5 Claude 4 Sonnet 20250514 74.60
Compare models. Tick up to 3 models in the table below, then run the comparison: who wins on which use, by how much, and how solid the measurement is.
| # | Model | Vendor | Score | Assessment | |
|---|---|---|---|---|---|
| 01 | — | 79.21 / 100 | Partial | ||
| 02 | Perplexity | 79.20 / 100 | Partial | ||
| 03 | ByteDance | 78.81 / 100 | Partial | ||
| 04 | Openhands | 77.61 / 100 | Partial | ||
| 05 | Atlassian | 76.81 / 100 | Partial | ||
| 06 | AI/Run Developer Agent | Epam systems, inc. | 76.80 / 100 | Partial | |
| 07 | ACoder | Acoder | 76.41 / 100 | Partial | |
| 08 | Harness AI | Harness | 74.81 / 100 | Partial | |
| 09 | JoyCode | Joycode | 74.61 / 100 | Partial | |
| 10 | Lingxi | 74.60 / 100 | Partial | ||
| 11 | Prometheus | — | 74.41 / 100 | Partial | |
| 12 | Refact.ai Agent | — | 74.40 / 100 | Partial | |
| 13 | Warp | Warp | 73.31 / 100 | Partial | |
| 14 | Anthropic Tools | — | 73.21 / 100 | Partial | |
| 15 | Salesforce SAGE | Salesforce ai research | 73.01 / 100 | Partial | |
| 16 | Anthropic | 71.71 / 100 | Partial | ||
| 17 | Anthropic | 71.70 / 100 | Partial | ||
| 18 | Qodo Command | Qodo | 71.21 / 100 | Partial | |
| 19 | Bloop | — | 71.20 / 100 | Partial | |
| 20 | Augment Agent | — | 70.41 / 100 | Partial | |
| 21 | Zencoder | — | 70.01 / 100 | Partial | |
| 22 | Google deepmind | 69.41 / 100 | Partial | ||
| 23 | OpenAI | 68.61 / 100 | Partial | ||
| 24 | — | 68.21 / 100 | Partial | ||
| 25 | Bytedance devinfra | 66.41 / 100 | Partial | ||
| 26 | Amazon Q Developer Agent | Aws | 65.41 / 100 | Partial | |
| 27 | PatchPilot | — | 64.61 / 100 | Partial | |
| 28 | W&B Programmer | — | 64.60 / 100 | Partial | |
| 29 | DeepSeek | 63.46 / 100 | Partial | ||
| 30 | Alibaba | 63.41 / 100 | Partial | ||
| 31 | NVIDIA | 63.21 / 100 | Partial | ||
| 32 | CodeStory Midwit | — | 62.21 / 100 | Partial | |
| 33 | OpenAI | 62.01 / 100 | Partial | ||
| 34 | ExpeRepair | — | 60.31 / 100 | Partial | |
| 35 | 60.21 / 100 | Partial | |||
| 36 | DeepSWE | — | 58.81 / 100 | Partial | |
| 37 | KGCompass | — | 58.31 / 100 | Partial | |
| 38 | devlo | — | 57.48 / 100 | Partial | |
| 39 | Emergent E1 | — | 57.21 / 100 | Partial | |
| 40 | Artemis Agent | — | 57.01 / 100 | Partial | |
| 41 | SWE-Rizzo | — | 56.61 / 100 | Partial | |
| 42 | Blackbox AI Agent | — | 55.91 / 100 | Partial | |
| 43 | FrogBoss | — | 53.61 / 100 | Partial | |
| 44 | Bracket.sh | — | 53.21 / 100 | Partial | |
| 45 | Gru | — | 52.86 / 100 | Partial | |
| 46 | 52.21 / 100 | Partial | |||
| 47 | EntroPO | — | 51.83 / 100 | Partial | |
| 48 | Engine Labs | — | 51.81 / 100 | Partial | |
| 49 | SemAgent | — | 51.71 / 100 | Partial | |
| 50 | AutoCodeRover | — | 51.61 / 100 | Partial |
Indicative score, weighted from public sources. Transparent methodology. See the AI ranking page for the full methodology and the other categories.
AI coding agent ranking
Le Recul's Coding agents ranking assesses systems that do not merely generate code but act on a project: Claude Code, Cursor Agent, Windsurf/Cascade, Aider and others. These agents read a code base, modify files, run tests and fix their own errors in a loop. They are therefore judged on their ability to carry a development task from end to end, not on the quality of an isolated snippet.
This is a distinct category from the Code ranking, which assesses raw models (the LLMs). Here the LLM is only one component: orchestration, tools, context management and the correction strategy count as much as the underlying model. Two agents built on the same model can produce very different results.
The sector moves fast and demonstrations are often flattering. Le Recul tries to measure what holds in real conditions — modifying a real repository, not breaking what exists, checking its own work — rather than the smoothness of a launch video. Read it as a critical reference point for choosing an agent according to your workflow, your language and your tolerance for risk.
What this ranking is for
This ranking helps you choose a development agent according to your real workflow: editor integration, autonomy, ability to test and repair, handling of large repositories.
It is a reminder that the same model gives different results depending on the agent orchestrating it: the complete system is what is assessed, not just the LLM.
How to read the ranking
- Ability to modify a real repository without breaking what exists.
- Execution and testing: does the agent check its own work?
- Correction loop: recovery after an error.
- Context handling on projects of realistic size.
- Supervision required and tolerance for risk.
FAQ — Coding agents
What is the difference with the Code ranking?
The Code ranking assesses raw models (LLMs). Coding agents add tools, execution, tests and a correction loop. The complete system is what is judged, not just the model.
Are two agents on the same model equivalent?
No. Orchestration, context handling and the correction strategy make a large difference, even with an identical model.
Can a coding agent work without supervision?
Not reliably. The best ones reduce the effort, but human review and testing remain necessary, especially on production code.
What is the assessment based on?
On the ability to carry real development tasks from end to end — modify, test, repair — following a documented method, rather than on scripted demonstrations.
Are open-source agents included?
Yes, when they are sufficiently documented. Some also belong to the Local AI views, for systems that can be run locally.