Design A2A Agent Card

Standards-compliant A2A Agent Card → advertises identity, skills, auth, caps for discovery.

Use When

• Discoverable by other A2A agents
• Expose caps → multi-agent orchestration
• Migrate agent → A2A protocol
• Public contract before impl
• Integrate → registries

In

• Required: Agent name + desc
• Required: Skills list (name, desc, I/O schemas)
• Required: Base URL host
• Optional: Auth (none, oauth2, oidc, api-key)
• Optional: Content types beyond text/plain
• Optional: Cap flags (streaming, push, state history)
• Optional: Provider org + URL

Do

Step 1: Identity + desc

1.1. Identity fields:

{
  "name": "data-analysis-agent",
  "description": "Performs statistical analysis, data visualization, and report generation on tabular datasets.",
  "url": "https://agent.example.com",
  "provider": {
    "organization": "Example Corp",
    "url": "https://example.com"
  },
  "version": "1.0.0"
}

1.2. Desc answers:

• Domains?
• Tasks?
• Limitations?

1.3. Canonical URL → /.well-known/agent.json.

→ Complete identity: name, desc, URL, provider, ver.

If err: Multi-domain → one agent w/ many skills vs many focused agents. A2A favors focused w/ clear bounds.

Step 2: Skills + I/O schemas

2.1. Define each:

{
  "skills": [
    {
      "id": "analyze-dataset",
      "name": "Analyze Dataset",
      "description": "Run descriptive statistics, correlation analysis, or hypothesis tests on a CSV dataset.",
      "tags": ["statistics", "data-analysis", "csv"],
      "examples": [
        "Analyze the correlation between columns A and B in my dataset",
        "Run a t-test comparing group 1 and group 2"
      ],
      "inputModes": ["text/plain", "application/json"],
      "outputModes": ["text/plain", "application/json", "image/png"]
    },
    {
      "id": "generate-chart",
      "name": "Generate Chart",
      "description": "Create bar, line, scatter, or histogram charts from tabular data.",
      "tags": ["visualization", "charts"],
      "examples": [
        "Create a scatter plot of height vs weight",
        "Generate a histogram of the age column"
      ],
      "inputModes": ["text/plain", "application/json"],
      "outputModes": ["image/png", "image/svg+xml"]
    }
  ]
}

2.2. Each skill:

• id: Unique (kebab-case)
• name: Display
• description: 1-2 sentences
• tags: Searchable
• examples: NL task examples
• inputModes: MIME accepts
• outputModes: MIME produces

2.3. Bounds clear + non-overlap. Each task → one skill.

→ Skills array w/ id, name, desc, tags, examples, I/O.

If err: Skills overlap → merge broader w/ more examples. Too broad → split focused.

Step 3: Auth

3.1. Scheme by deploy:

No auth (local/trusted):

{
  "authentication": {
    "schemes": []
  }
}

OAuth 2.0 (rec prod):

{
  "authentication": {
    "schemes": ["oauth2"],
    "credentials": {
      "oauth2": {
        "authorizationUrl": "https://auth.example.com/authorize",
        "tokenUrl": "https://auth.example.com/token",
        "scopes": {
          "agent:invoke": "Invoke agent skills",
          "agent:read": "Read task status"
        }
      }
    }
  }
}

API Key (shared-secret):

{
  "authentication": {
    "schemes": ["apiKey"],
    "credentials": {
      "apiKey": {
        "headerName": "X-API-Key"
      }
    }
  }
}

3.2. Min viable:

• Local dev → none
• Internal svc → apiKey
• Public → oauth2 / oidc

3.3. Doc token/key provisioning → provider section or external docs.

→ Auth block matches deploy sec reqs.

If err: No OAuth infra → start apiKey + plan migration. NEVER public w/ none.

Step 4: Caps

4.1. Protocol features:

{
  "capabilities": {
    "streaming": true,
    "pushNotifications": false,
    "stateTransitionHistory": true
  }
}

4.2. Flags by impl:

• streaming: true if SSE via tasks/sendSubscribe. Real-time progress.
• pushNotifications: true if webhook callbacks on state. Requires store + call webhooks.
• stateTransitionHistory: true if full history (submitted, working, completed). Audit.

4.3. Only true if fully supported. Advertising unsupported → breaks interop.

→ Caps obj w/ flags matching impl.

If err: Unsure → false. Add in future. Removing = breaking change.

Step 5: Validate + publish

5.1. Assemble:

{
  "name": "data-analysis-agent",
  "description": "Performs statistical analysis and visualization on tabular datasets.",
  "url": "https://agent.example.com",
  "version": "1.0.0",
  "provider": {
    "organization": "Example Corp",
    "url": "https://example.com"
  },
  "authentication": {
    "schemes": ["oauth2"],
    "credentials": { ... }
  },
  "capabilities": {
    "streaming": true,
    "pushNotifications": false,
    "stateTransitionHistory": true
  },
  "skills": [ ... ],
  "defaultInputModes": ["text/plain"],
  "defaultOutputModes": ["text/plain"]
}

5.2. Validate:

• Parse JSON, no syntax errs
• Required fields (name, desc, url, skills)
• Each skill: id, name, desc, ≥1 I/O mode
• URL reachable, card at /.well-known/agent.json

5.3. Publish:

• Serve → https://<agent-url>/.well-known/agent.json
• Content-Type: application/json
• CORS if cross-origin
• Register → agent directories

5.4. Test discovery:

curl -s https://agent.example.com/.well-known/agent.json | python3 -m json.tool

→ Valid JSON served at well-known URL, parseable by A2A client.

If err: JSON fail → lint. URL unreachable → DNS, SSL, web server. CORS → Access-Control-Allow-Origin.

Check

• Valid JSON, no syntax errs
• Required: name, desc, url, skills
• Each skill: id, name, desc, inputModes, outputModes
• Auth matches deploy sec
• Caps reflect impl
• Served at /.well-known/agent.json + Content-Type
• A2A clients fetch + parse OK
• Examples realistic + trigger correct skill

Traps

• Overpromising caps: streaming: true / pushNotifications: true w/o impl → client fails. Conservative.
• Vague skill desc: "does data stuff" → no match. Specific I/O + domains.
• Missing CORS: Browser A2A clients can't fetch w/o CORS.
• Skill overlap: 2 skills same task → clients can't pick. Clear bounds.
• Forget default modes: Missing defaultInputModes/defaultOutputModes → clients no MIME.
• Ver stagnation: Update ver when skills/caps change. Clients cache.
• Publish before impl: Card = contract. Publishing un-impl'd → runtime fails.

→

• implement-a2a-server — server behind card
• test-a2a-interop — validate conformance + interop
• build-custom-mcp-server — MCP alt/complement
• configure-mcp-server — MCP patterns applicable