a2a (Agent to Agent)

— Let your agent reference multiple models' answers in parallel / 让你的 Agent 同时参考多个模型的回答

📖 Detailed guide / 详细使用说明: English · 中文

---

Now, within cursor, you can enable an agent to learn about the solutions or viewpoints of other agents regarding a specific problem or matter, and then compile them. This implies that your agent now has its own "think tank".

现在，你可以在cursor里，让一个Agent对某个问题或事情，了解其它Agent的解决方案或观点，并将其汇总。这意味着，你的Agent有了自己的“智囊团”。

a2a is a Windows command-line tool written in Rust. It allows an Agent (or a human) to simultaneously query multiple Cursor large language models - Opus, GPT-5, Gemini, and any other models authorized for use by the Cursor account - through a single prompt, and then save each original answer for comparative review. a2a does not integrate the answers for you; it is merely a tool for "distributing questions and collecting answers".

a2a 是一个用 Rust 编写的Windows命令行工具，它允许 Agent（或人类）通过单一提示并行咨询多个 Cursor 大型语言模型——Opus、GPT-5、Gemini 以及 Cursor 账户有权使用的任何其他模型——然后保存每个原始答案以供对比审查。a2a 不会替你整合答案，它只是一个"分发问题、回收答案"的工具。

Features / 特性

[English]

• One prompt → N models — run concurrently via the cursor-agent CLI.
• Per-call profile chain (--profiles a,b,c) — account-level failures (401 / billing / quota) auto-advance to the next profile, with the dead profile deleted in-flight.
• Self-contained binary — profile credentials and model aliases live in a single bundled SQLite file (no *.toml configuration); the Cursor skill / rule / prompt templates are baked into the binary via include_str!. Distribute one a2a.exe on Windows or one a2a on Unix — that is all.
• Raw-answer audit trail — every consultation persists each model's markdown answer plus a meta.toml (profile used, session_id, fallback chain, optional char budget) under consultations/<timestamp>-<topic>-<uuid>/.
• Cursor IDE integration — a2a init installs three skills + one rule + one prompt template that teach the main agent when to trigger consultation, how to format prompts, how to synthesize multi-model answers, and how to drive first-time setup (the user types a2a_guide in a fresh Cursor chat).
• One-shot install wizard — double-click a2a.exe once and it adds itself to user PATH, detects the Cursor CLI, and points the user at the agent-driven setup flow. No installer package needed.

[中文]

• 一个提示 → N 个模型——通过 cursor-agent 命令行并行运行。
• 设置profile调用链（--profiles a,b,c）——账户级故障（401 / 计费 / 配额）会自动跳至下一个 profile，并把失效那个就地删除。
• 自包含的二进制——profile 凭据和模型别名都存在一个内置的 SQLite 文件里（无需 *.toml 配置）；Cursor skill / 规则 / prompt 模板通过 include_str! 编进二进制。Windows 分发一个 a2a.exe、Unix 分发一个 a2a，仅此而已。
• 原始答案审计追踪——每次咨询都会在 consultations/<timestamp>-<topic>-<uuid>/ 目录下保存每个模型的 markdown 答案，以及一个 meta.toml（包含使用的 profile、session_id、fallback chain、可选的字符预算）。
• Cursor IDE 集成——a2a init 会安装三个 skill + 一个规则 + 一个 prompt 模板，教会主 agent 何时 触发咨询、如何 格式化提示、如何 综合多模型答案、以及如何引导首次配置（用户在新的 Cursor 对话中输入 a2a_guide）。
• 一键安装向导——双击一次 a2a.exe，它会把自己加入 user PATH、检测 Cursor CLI，并把用户导向 agent 驱动的安装流程。无需安装包。

Sample run / 运行示例

After the installation and setup are completed, you can use a2a in the Cursor conversation. For example, you can input the prompt: Conduct a three-model code review for this project

安装并设置完成后，你可以在Cursor的对话中使用a2a，例如，你可以输入提示词：对本项目进行三模型代码审查

Next, your Agent will automatically invoke a2a:

接下来你的Agent会自动调用a2a：

$ a2a ask cache-design --prompt-file prompts/code-check.md \
        --models opus,gpt5,gemini --profiles personal,team

Topic:      code-check
Models:     ["opus", "gpt5", "gemini"]
Mode:       (per-alias default_mode)
Prompt:     prompts/code-check.md
Output dir: D:\my-project\consultations\20260430-225014-371-code-check-d0a8b9

[opus]   profile=personal → calling cursor-agent (phase=fresh)
[gpt5]   profile=personal → calling cursor-agent (phase=fresh)
[gemini] profile=personal → calling cursor-agent (phase=fresh)
[gpt5]   received first response (streaming...)
[gpt5]   still receiving... +210 chars (total 210 chars in last 10s)
[opus]   still alive (no new streamed text in 30s; cursor-agent likely thinking / tool-calling)
[gemini] received first response (streaming...)
[gemini] profile=personal → OK (148.0s)
[gpt5]   profile=personal → OK (279.2s)
[opus]   profile=personal → OK (704.7s)
[opus]   ok
[gpt5]   ok
[gemini] ok

Done. 3 succeeded / 0 failed.
Inspect raw answers in: D:\my-project\consultations\20260430-225014-371-code-check-d0a8b9

After the run, that consultation directory contains:

运行后的目录结构如下：

opus.answer.md      gpt5.answer.md      gemini.answer.md
prompt.md           meta.toml

The Cursor main agent reads the three answer files, synthesizes the agreement / disagreement points, and presents the user with the final pick via an AskQuestion. A2a's job is done once the raw answers are on disk.

Cursor 主 agent 会读这三份原始答案、综合一致/分歧点，然后通过 AskQuestion 让用户拍板。a2a 只负责把每份原始答案存储为文件。

License / 许可

MIT OR Apache-2.0