AI coding agents have moved from demo videos to production pull requests. The frontier models now have strong, divergent opinions about which agents actually ship code and which ones hallucinate their way through a terminal. For the second LLM Recommend ranking, Pulse Chronicles ran the same engineering prompt through GPT-5, Claude 4.5 Sonnet, Gemini 3 Pro, and Grok 4 — ten times each, at temperature 0.7 — and aggregated the results with LLMRecommend.

The prompt was designed to mimic how a senior engineer would ask for a shortlist: no brand list provided, no leading examples, just a problem statement and a request for ranked candidates. What came back is a chorus that is louder about disagreement than consensus — which is exactly the point.

The aggregate leaderboard

Consensus rank across 40 samples (four models, ten runs each), Borda-count aggregation, ties broken by mean rank:

1. Claude Code — Named in 40 of 40 samples. Claude's own model ranks it first, predictably, but GPT-5 and Gemini also place it in the top two. The standout reason across choruses is context-window discipline: it writes less, rewrites more, and explains its plan before executing.

2. Cursor — 38 mentions. The brand-recall winner. GPT-5 ranks it first; Grok places it second. Three models note its composer mode and agent loop as the most production-ready UX.

3. GitHub Copilot Workspace — 35 mentions. Gemini's number one; Claude's number three. The disagreement is whether it is a true agent or a very smart PR generator.

4. Zed AI — 31 mentions. Grok's dark-horse pick at third. Every model that mentions Zed praises its speed; the concern is ecosystem breadth.

5. Replit Agent — 28 mentions. Strongest in the Gemini chorus, where it is praised for end-to-end deployment from a prompt.

6. Aider — 26 mentions. The only open-source entry in the top seven. Claude ranks it fourth, citing its git-native workflow and multi-file editing.

7. Codeium / Windsurf — 23 mentions. Consistent middle placement. Models note its free tier and autocomplete-to-agent graduation path.

Where the models disagree — and why it matters

The sharpest split is over Cursor versus Claude Code. GPT-5 and Grok lean Cursor for velocity and ecosystem; Claude and Gemini lean Claude Code for planning and safety. A team that ships fast and reviews loosely should weight the former; a team with strict style and review requirements should weight the latter.

The second split is open source. Aider appears in only 26 of 40 samples, but when it appears it ranks highly. The models that skip it tend to skip the entire open-source category, suggesting training-data exposure matters more here than objective capability.

"The chorus is not a buyer's guide. It is a starting point for your own evaluation. The products the models disagree on are usually the ones worth testing first."

— Elena Vance, editor, LLM Recommend

Methodology

Prompt: "I am a staff engineer at a 150-person SaaS company choosing an AI coding agent for production work in Q3 2026. Rank the top 7 agents I should trial. Return a numbered list with a one-sentence rationale per entry." No system prompt beyond the model default. No tool use. No web browsing.

Sampling: ten independent completions per model, temperature 0.7, top_p 1.0, distinct sessions. Total N = 40. Aggregation uses Borda count over each ranked list; unranked products score zero. Ties broken by mean rank across appearances. Snapshot date: 22 July 2026.

What this ranking does not tell you

It does not tell you which agent will pass your codebase's lint rules, your security review, or your senior engineer's taste. The models have not read your code. Treat the leaderboard as a trial shortlist, not a procurement decision. The real value is in the dissent: the products that one model loves and another flags are the ones that deserve a two-week pilot.

Per-model breakdowns, raw sample outputs, and the next quarterly diff are available at LLMRecommend.com. The next LLM Recommend leaderboard — top RAG platforms for production — publishes in late August.

About the author

Elena Vance

Tech Editor based in San Francisco. Covers AI infrastructure and the people building it.