AI ranking
AI agents
General-purpose agentic systems (OpenClaw, Claude Cowork, Manus, OpenClaude, Genspark, Devin-like). Assessed on real long-running tasks, not as pure LLMs.
Visual top 10 — Le Recul score
- 01 OpenClaw 87.30
- 02 OpenHands 87.30
- 03 OpenClaude 85.50
- 04 ChatGPT Agent 76.50
- 05 Microsoft Copilot Cowork 76.50
- 06 Gemini Enterprise 73.80
- 07 SWE-agent 69.30
- 08 Aider Architect 67.50
- 09 OpenCode 67.50
- 10 Hermes Agent 65.70
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 | Open-source agent by Peter Steinberger for browser automation with the LLM of your choice. | Peter steinberger | 87.30 / 100 | Le Recul | |
| 02 | Open-source coding agent (formerly OpenDevin), bring your own model. Present in SWE-bench. | All hands ai | 87.30 / 100 | Le Recul | |
| 03 | Open-source agent inspired by Claude Cowork, bring your own model. | Anthropic community | 85.50 / 100 | Le Recul | |
| 04 | OpenAI agent built into ChatGPT, able to browse the web and handle files. | OpenAI | 76.50 / 100 | Le Recul | |
| 05 | Cowork variant of Microsoft Copilot for multi-agent collaboration. | Microsoft | 76.50 / 100 | Le Recul | |
| 06 | Google enterprise offering built on Gemini, with customisable agents. | 73.80 / 100 | Le Recul | ||
| 07 | Academic agentic scaffold for SWE-bench, bring your own model. | Princeton nlp | 69.30 / 100 | Le Recul | |
| 08 | Aider's architect mode — agentic planning separated from editing. Bring your own model. | Aider | 67.50 / 100 | Le Recul | |
| 09 | Open-source command-line coding agent, an alternative to Claude Code. | Opencode community | 67.50 / 100 | Le Recul | |
| 10 | Nous Research agent built on Hermes 3, open source, bring your own model. | Nous research | 65.70 / 100 | Le Recul | |
| 11 | Anthropic command-line coding agent, built on Claude. | Anthropic | 64.80 / 100 | Le Recul | |
| 12 | OpenAI command line for Codex/GPT, built into the OpenAI ecosystem. | OpenAI | 64.80 / 100 | Le Recul | |
| 13 | GitHub Copilot command line, built into GitHub. | Github / microsoft | 64.80 / 100 | Le Recul | |
| 14 | Anthropic collaborative workspace built on Claude — email, docs, research, analysis. | Anthropic | 62.10 / 100 | Le Recul | |
| 15 | Web research and synthesis agent, an alternative to Perplexity. | Genspark ai | 61.20 / 100 | Le Recul | |
| 16 | Autonomous multi-task agent positioned as a general-purpose alternative to cloud assistants. | Manus ai | 60.30 / 100 | Le Recul | |
| 17 | Google agentic studio built into Workspace (Gmail, Docs, Sheets). | 58.50 / 100 | Le Recul | ||
| 18 | Microsoft agentic suite built into Microsoft 365, Outlook and Teams. | Microsoft | 58.50 / 100 | Le Recul | |
| 19 | Droid Factory.ai agentic platform for software development. | Factory.ai | 54.90 / 100 | Le Recul |
Indicative score, weighted from public sources. Transparent methodology. See the AI ranking page for the full methodology and the other categories.
General-purpose AI agent ranking
Le Recul's AI agents ranking assesses general-purpose agentic systems — OpenClaw, Claude Cowork, Manus, OpenClaude, Genspark and other autonomous assistants — on long, real tasks, and not as plain language models. An agent is not reducible to its LLM: it chains actions, uses tools, observes the result and corrects itself. It is that ability to see a task through that is measured here.
The field is young and noisy. Many impressive demonstrations do not reproduce in real conditions, and public APIs are often missing to assess these systems consistently. Le Recul therefore relies on an editorial catalogue scored by its own method (autonomy, capability, reliability), with an acknowledged editorial ceiling: no agent is presented as magic.
This ranking measures a maturity, not a promise. It separates general-purpose agents from specialised ones — coding agents, browser agents, business agents — which have their own categories. Read it as a critical state of play of a sector in motion, useful for understanding what these systems can really do today, and where they still fail.
What this ranking is for
This ranking helps you tell the agents genuinely able to carry a long task through from those that impress in a demo but drop out in real conditions.
It separates general-purpose agents from specialised ones (code, browser, business), to avoid comparing systems that are not aiming at the same use.
How to read the ranking
- Autonomy: chaining steps without constant human takeover.
- Real capability on long tasks, not on scripted demos.
- Reliability and reproducibility of the result.
- Tool use and verification of the output produced.
- Le Recul editorial score (acknowledged ceiling), for want of consistent public APIs.
FAQ — AI agents
Why a « Le Recul » score and not a standard benchmark?
Because there are not yet consistent public APIs to measure these agents. Le Recul uses an editorial catalogue scored on autonomy, capability and reliability, with an acknowledged ceiling.
What is the difference with the Chat ranking?
Chat assesses language models in conversation. Agents chain actions, use tools and correct themselves: they are judged on long tasks, not on a single answer.
Why are some agents not listed?
For want of enough data to assess them seriously. Systems that are too thinly documented are tracked separately rather than overrated.
Will a well-ranked agent do my work on its own?
No. Even the best require supervision and verification. The ranking measures relative maturity, not full autonomy.
Where are the coding and browser agents?
In their dedicated categories: Coding agents and Browser agents. This page covers general-purpose agents.