How AgentRank Scores and Ranks Mainstream Agents
AgentRank is a review catalog of mainstream Agent products: what they can do, how they feel, and a composite score from this platform. It is not a package marketplace, and it is not a third-party lab bench. The scores below are AgentRank catalog scores, taken from the catalog snapshot used to build the site.
The method is the one already published on the About page, and this post does not add weights or a formula that page does not give.
How scoring works
Each product is stored as bilingual markdown. The composite score (0–100) blends four dimensions:
- Capability
- Reliability
- Ecosystem
- Value
About describes those four as the blend that produces the composite score. It does not publish a numeric weight for each dimension, so this article does not invent one.
Editorial rank is a separate number. It is the recommended order on the homepage, not a sort of the composite score. That distinction is visible in the current catalog: GitHub Copilot has a composite score of 90, and ChatGPT Agent has a composite score of 89, but ChatGPT Agent is rank 3 and Copilot is rank 4. A higher score does not automatically mean a higher recommended place on the homepage.
The shortlist exists so you can compare coding IDEs, autonomous workers, and research agents without reading every landing page. Categories in the catalog include coding, general, autonomous, research, and open source. Individual product pages such as Cursor, Claude Code, Devin, and Perplexity are the place to read the review, the four-dimension breakdown, and the official product link.
The same catalog is also available to machines over REST and JSON-RPC at /api/mcp, including initialize, tools/list, and tools/call. That is an access path, not a second scoring method.
Current catalog scores
These figures are AgentRank catalog values (each entry’s lastUpdated in the snapshot is 2026-11-04). Rank is editorial homepage order. Score is the composite 0–100.
| Name | Rank | Score |
|---|---|---|
| Cursor | 1 | 94 |
| Claude Code | 2 | 93 |
| ChatGPT Agent | 3 | 89 |
| GitHub Copilot | 4 | 90 |
| Devin | 5 | 86 |
| Windsurf | 6 | 85 |
| Manus | 7 | 84 |
| Gemini | 8 | 83 |
| Cline | 9 | 82 |
| Trae | 10 | 81 |
| Perplexity | 11 | 80 |
| Replit Agent | 12 | 78 |
Read the table as a shortlist, not as a universal “best tools” list. The catalog already says, for example, that Cursor is the default pick for coding agents, that Perplexity is a research layer rather than a coding agent, and that Replit Agent is a fit for prototypes rather than a serious brownfield codebase. Those judgements live on the product pages, next to the numbers.
How to use a score
Start with the job, then open the matching product page.
If you want a coding IDE agent, the catalog’s current default pick is Cursor (rank 1, score 94). Claude Code sits next to it (rank 2, score 93) as a CLI-first option. A longer comparison that uses only these catalog scores is in Cursor vs Claude Code, using AgentRank scores.
If you want a GitHub-native assistant, open GitHub Copilot. If you want a queued, issue-to-PR worker, open Devin. If you want cited research rather than repo edits, open Perplexity. The composite score is a way to compare entries in this catalog. It is not a lab measurement from an outside benchmark, and it is not a claim about packages you can install from a marketplace.
When the homepage order and the composite score disagree, trust the About page’s wording: editorial rank is the recommended order; the score is the blend of the four dimensions. Use both, and read the review on the agent page before you treat either number as a decision.