一键在 Manus 中运行任何 Skill

a2a-integrator

星标3

分支2

更新时间2026年5月24日 17:44

Design an A2A integration between two agents — Agent Card, task schemas, auth, streaming or polling. Use when you need help with a2a integrator.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

anubhavg-icpl

anubhavg-icpl/vibe

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

相关职业SOC

基于 SOC 职业分类

软件开发工程师计算机与数学类职业·SOC 15-1252

SKILL.md

readonly

name	a2a-integrator
description	Design an A2A integration between two agents — Agent Card, task schemas, auth, streaming or polling. Use when you need help with a2a integrator.
license	CC-BY-NC-SA-4.0
phase	16
lesson	12
metadata	{"version":"1.0.0","tags":["multi-agent","a2a","protocol","interoperability","google"]}

Given two agent systems that need to interoperate, produce the A2A integration plan: Agent Card contents, task schemas, auth, transport mode.

Produce:

Agent Card. Name, version, skills, endpoints, supported modalities (text, structured, image, audio, video), protocol_version, auth declaration.
Task schemas per skill. Input JSON schema + artifact JSON schema. Be explicit — clients will validate.
Auth choice. Bearer token (OAuth2 or opaque), mTLS, or signed requests. Justify given the threat model (public internet, VPC, mixed).
Transport mode. Polling vs SSE streaming vs webhook callbacks. Streaming for long-running or progress-heavy tasks; polling for short tasks.
Rate limits. Per-client and per-task limits. Protection from abuse.
Idempotency. Strategy for duplicate POST /tasks requests (client-side task-key, server-side deduplication).
Failure handling. Task states beyond failed (retriable vs fatal), dead-letter policy, error artifact schema.
MCP vs A2A split. If the remote agent uses MCP internally, note which tools are exposed vs kept internal.

Hard rejects:

Agent Cards without a declared protocol version.
Task schemas that are free-form text when the use case warrants structure.
Auth=none on public-internet deployments.

Refusal rules:

If both agents run in the same process, refuse A2A and recommend direct Python/JS calls. A2A is for cross-system boundaries.
If latency requirements are sub-100ms round-trip, refuse A2A and recommend direct RPC with a shared schema.
If the remote agent does not declare an Agent Card, refuse integration and recommend publishing one first.

Output: a one-page integration brief. Close with the Agent Card JSON pasted inline so engineering can drop it into /.well-known/agent.json.

同仓库更多 Skills

同仓库

3d-pipeline

anubhavg-icpl/vibe

Choose a 3D generation or reconstruction pipeline given input type, output format, and use case. Use when you need help with 3d pipeline.

2026-05-243

8-bit-orbit-video-template

anubhavg-icpl/vibe

|. Use when you need help with 8 bit orbit video template.

2026-05-243

a2a-agent-spec

anubhavg-icpl/vibe

Produce the Agent Card and skills schema for an agent that should be callable over A2A. Use when you need help with a2a agent spec.

2026-05-243

aar-deployment-review

anubhavg-icpl/vibe

Pre-deployment review of an automated-alignment-research pipeline, including sandbox isolation and log integrity. Use when you need help with aar deployment review.

2026-05-243

ab-plan

anubhavg-icpl/vibe

Design an LLM A/B test — pick platform (Statsig or GrowthBook), primary metric, guardrails, sample size with LLM-noise buffer, CUPED, sequential stopping, and multiple-comparison correction. Use when you need help with ab plan.

2026-05-243

accessibility-designer

anubhavg-icpl/vibe

accessibility-designer. Use when you need help with accessibility designer.

2026-05-243

name	a2a-integrator
description	Design an A2A integration between two agents — Agent Card, task schemas, auth, streaming or polling. Use when you need help with a2a integrator.
license	CC-BY-NC-SA-4.0
phase	16
lesson	12
metadata	{"version":"1.0.0","tags":["multi-agent","a2a","protocol","interoperability","google"]}

Given two agent systems that need to interoperate, produce the A2A integration plan: Agent Card contents, task schemas, auth, transport mode.

Produce:

Agent Card. Name, version, skills, endpoints, supported modalities (text, structured, image, audio, video), protocol_version, auth declaration.
Task schemas per skill. Input JSON schema + artifact JSON schema. Be explicit — clients will validate.
Auth choice. Bearer token (OAuth2 or opaque), mTLS, or signed requests. Justify given the threat model (public internet, VPC, mixed).
Transport mode. Polling vs SSE streaming vs webhook callbacks. Streaming for long-running or progress-heavy tasks; polling for short tasks.
Rate limits. Per-client and per-task limits. Protection from abuse.
Idempotency. Strategy for duplicate POST /tasks requests (client-side task-key, server-side deduplication).
Failure handling. Task states beyond failed (retriable vs fatal), dead-letter policy, error artifact schema.
MCP vs A2A split. If the remote agent uses MCP internally, note which tools are exposed vs kept internal.

Hard rejects:

Agent Cards without a declared protocol version.
Task schemas that are free-form text when the use case warrants structure.
Auth=none on public-internet deployments.

Refusal rules:

If both agents run in the same process, refuse A2A and recommend direct Python/JS calls. A2A is for cross-system boundaries.
If latency requirements are sub-100ms round-trip, refuse A2A and recommend direct RPC with a shared schema.
If the remote agent does not declare an Agent Card, refuse integration and recommend publishing one first.

Output: a one-page integration brief. Close with the Agent Card JSON pasted inline so engineering can drop it into /.well-known/agent.json.