Written by: Mariana Fonseca, Editorial Team, AI Growth Agent
Key Takeaways
- Agent cards are structured JSON declarations at /.well-known/agent-card.json that make sites discoverable and actionable to autonomous AI agents.
- Four card types, Decision Tiles, Rich Data Summaries, Purchase Workflows, and Dynamic Feeds, cover the full range of headless marketing use cases.
- Agent cards connect with MCP servers, llms.txt files, and the A2A protocol so agents can both discover capabilities and execute them.
- Independent implementations on any headless stack deliver the same A2A-compliant functionality without Salesforce licensing or per-conversation fees.
- AI Growth Agent provisions complete agent cards, MCP endpoints, and llms.txt files on your domain within the first week, and you can schedule a demo to get started.
Why Agent Cards Matter More Than Traditional Pages
The shift to agentic discovery accelerated sharply in 2026. Google’s Agent2Agent (A2A) protocol, originally introduced in 2025 and brought to production-grade v1.0 in early 2026, is now governed by the Linux Foundation and supported by more than 150 organizations including Microsoft, AWS, Salesforce, SAP, and ServiceNow. It standardizes how agents advertise capabilities, exchange tasks, and communicate over the wire.
An agent does not scroll, which means it cannot see hero images, navigation bars, or brand taglines that human visitors rely on. Instead, it reads structured metadata, follows declared standards, and routes requests to endpoints it can call. Because a page that declares nothing exposes no machine-readable signals, the agent treats it as undocumented. Traditional CMS pages built for human readers carry no capability declarations, no machine-readable endpoint map, and no authentication schema. As a result, they are invisible to the agents now doing most of the discovery.
Agent cards fix this gap by giving every compliant agent a known address, a structured format, and a complete description of what the site can do. The canonical location is /.well-known/agent-card.json, and any A2A-compliant agent knows to look there first. These cards are organized into four distinct types, and each type maps to a specific category of agent interaction that headless marketing requires.
The Four Types of Agent Cards for Headless Marketing
Decision Tiles present a bounded set of choices an agent can surface directly inside a conversational interface. In headless marketing, decision tiles power upgrade prompts such as “choose your plan tier,” discount actions such as “apply a loyalty code,” and product configuration flows. The agent renders the tile natively in Slack, Microsoft Teams, or any MCP-compatible client, and it does this without opening a browser. Salesforce’s Agentforce Experience Layer demonstrated this pattern at TDX 2026. That launch separated agent behavior from presentation so a single experience definition renders across every surface under a “build once, render everywhere” model.
Rich Data Summaries return structured data objects an agent can read, cite, and act on. Campaign ROI dashboards, competitor traffic snapshots, and audience segment reports all fit this pattern. The card declares the endpoint, the schema of the response, and the authentication method. With that information in place, the agent can pull a live data object without a human intermediary.
Purchase Workflows expose multi-step transactional sequences as callable endpoints. Booking, event registration, and payment flows are the primary headless marketing use cases. The card declares each step as a capability. The agent then executes the sequence on the user’s behalf, and the transaction completes without a browser session. Google’s I/O 2026 keynote extended agentic booking to local services from home repair to pet care. That announcement confirmed that purchase workflows now sit at the level of a production-grade expectation.
Dynamic Feeds return personalized, real-time content streams. Product recommendation engines, personalized editorial feeds, and location-aware offer lists all map to this card type. The agent passes a user context object, the endpoint returns a ranked feed, and the agent surfaces the result inside whatever interface the user already uses.
Agent Card JSON Schema You Can Use Today
The following schema is copy-pasteable for any non-Salesforce headless stack. Fields follow the A2A protocol specification announced by Google in April 2025.
{ "schema_version": "a2a-v1.0", "name": "Brand Marketing Agent", /* Human-readable name surfaced in agent discovery UIs */ "description": "Headless marketing agent exposing campaign data, purchase workflows, and personalized content feeds.", "url": "https://yourdomain.com/.well-known/agent-card.json", /* Canonical self-referencing URL required by A2A spec */ "capabilities": [ { "type": "decision_tile", "id": "upgrade_prompt", "description": "Surfaces plan upgrade options with pricing context.", "endpoint": "https://yourdomain.com/api/agent/upgrade", "method": "POST" }, { "type": "rich_data_summary", "id": "campaign_roi", "description": "Returns campaign ROI data for a given date range.", "endpoint": "https://yourdomain.com/api/agent/roi", "method": "GET", "parameters": { "start_date": "ISO8601", "end_date": "ISO8601" } }, { "type": "purchase_workflow", "id": "event_registration", "description": "Executes a multi-step event registration and payment flow.", "endpoint": "https://yourdomain.com/api/agent/register", "method": "POST", "steps": ["select_event", "collect_attendee", "process_payment"] }, { "type": "dynamic_feed", "id": "personalized_recommendations", "description": "Returns a ranked content or product feed based on user context.", "endpoint": "https://yourdomain.com/api/agent/feed", "method": "POST", "parameters": { "user_context": "object", "limit": "integer" } } ], "authentication": { "type": "bearer", "token_endpoint": "https://yourdomain.com/api/auth/token" /* Replace with OAuth2 or API-key scheme as required */ }, "mcp_server": { "url": "https://yourdomain.com/mcp", "version": "2025-11-05" /* MCP server URL where tools are exposed for agent interaction */ }, "llms_txt": "https://yourdomain.com/llms.txt", /* Points LLM crawlers to the site's structured content index */ "contact": "[email protected]" }Every field in this schema carries a specific job. The capabilities array maps directly to the four card types described above. The mcp_server block tells the agent where to find the tool-access layer beneath the card. The llms_txt field points LLM crawlers to the structured content index. Leaving out any of these fields leaves the agent with an incomplete picture of what the site can do.
How Agent Cards, MCP, and llms.txt Work Together
Google’s developer guide on A2A draws a clear protocol boundary. MCP connects agents to tools and data, while A2A connects agents to other agents through agent cards. The two layers work together but do not replace each other. A site with an MCP server and no agent card behaves like a service running on an unpublished number. A site with an agent card and no MCP server declares capabilities it cannot fulfill.
The discovery sequence follows a predictable path. An A2A-compliant agent hits /.well-known/agent-card.json first. It reads the card, identifies the declared capabilities, and resolves the MCP server URL from the mcp_server block. It then connects to the MCP server to access the tools that fulfill those capabilities. Solace Agent Mesh documentation describes this pattern in production. In that model, agent cards are published as events on discovery topics, and orchestrators subscribe to capability announcements and route requests dynamically without hard-coded endpoints.
llms.txt and llms-full.txt serve a parallel function for LLM crawlers that read content rather than call endpoints. They provide a structured index of the site’s content, similar to robots.txt but designed for systems that run citation passes instead of traditional crawlers. Natural language query parameters via /?s={query} extend this pattern further. An agent can pass a query directly into the URL and receive a personalized, internally linked response without a separate API call. Pages served in Markdown to agent crawlers improve citation quality because LLMs parse Markdown more reliably than HTML.
Running Headless Marketing Without Salesforce
Salesforce Headless 360, announced April 15, 2026 at TDX, exposes the entire Salesforce platform through more than 60 MCP tools, over 30 preconfigured coding skills, and CLI commands. The architecture is technically sophisticated. It is also a CRM-first architecture that begins with unified customer data in Service Cloud and Data 360, then layers agentic capabilities on top. Salesforce Agentforce pricing starts at $2 per conversation, with Flex Credits at $0.10 per action, tied to the Salesforce platform. For brands not already running on Salesforce Customer 360, the integration, configuration, and data strategy overhead is substantial before a single agent card ships.
Independent implementations serve the same A2A standard without that platform dependency. AI Growth Agent provisions the full stack on the client’s own domain. That stack includes Blog MCP, agent discovery via /.well-known/, complete agent card JSON, llms.txt and llms-full.txt, natural language query parameters, Markdown serving for agent crawlers, and the reverse proxy rewrite that connects everything to a subdirectory under the brand’s domain. The client owns the site. No Salesforce license, no per-conversation billing, and no agency in the loop.
Deployment Checklist for Any Headless Engine
- Serve
/.well-known/agent-card.json. Place the complete JSON schema at the canonical A2A discovery URL. Validate that the file is publicly accessible and returns the correctContent-Type: application/jsonheader. - Expose MCP endpoints. Stand up an MCP server that fulfills the capabilities declared in the agent card. Confirm the
mcp_server.urlin the card resolves to a live, authenticated endpoint. - Publish
llms.txtandllms-full.txt. Place both files at the root of the domain.llms.txtprovides the summary index, andllms-full.txtprovides the complete structured content map for LLM crawlers. - Configure the reverse proxy. Route the blog and all agentic discovery files through a subdirectory on the brand’s domain, not a subdomain. Subdirectory placement passes domain authority to every indexed page, while a subdomain is treated as a separate site by Google.
- Enable bot tracking. Instrument per-page bot tracking to capture every crawl, citation, and training sweep. Without bot tracking, you cannot confirm that agents are finding and reading the declared capabilities.
- Enable natural language query parameters. Configure
/?s={query}to return personalized, internally linked responses so agents can query the site directly in natural language. - Serve Markdown to agent crawlers. Detect agent user agents and serve Markdown versions of content pages to improve LLM parse quality and citation accuracy.
- Validate with A2A-compliant agents. Run a discovery pass using an A2A-compliant agent to confirm the card is readable, the MCP tools are callable, and the authentication flow completes without errors.
Why Autonomous Engines Keep Agent Cards Accurate
Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. The same research estimates that more than 40% of agentic AI projects could be canceled by 2027 because of unclear value, rising costs, and weak governance. The gap between adoption intent and production deployment is real. Only about 11% of agentic AI pilots reach full production deployment due to integration, governance, and change-management friction.
Agent cards do not behave like a one-time configuration. The A2A standard evolves, MCP tool schemas change, and new capability types emerge. Content that feeds rich data summaries and dynamic feeds goes stale as soon as the underlying data changes. A static agent card deployed once and never updated becomes a liability. It declares capabilities the site no longer fulfills, which breaks agent trust and removes the site from agentic discovery flows.
AI Growth Agent runs weekly universe refreshes across 3,000+ searches, self-heals content as the world changes, and rolls out new agentic standards across every client site within the same rapid deployment window described earlier. The memory system enforces brand rules without human intervention. When Google’s Web MCP standard released, AI Growth Agent had already been running Blog MCP in production for roughly a year, with clients live in the summer of 2025. That lead time often decides whether an agent finds your site first or ignores it when the next wave of agents rolls out.
Frequently Asked Questions
What is the difference between an agent card and a robots.txt file?
Robots.txt tells traditional crawlers where to look and where to stay out. It was never designed to describe an MCP server, declare an agent endpoint, or carry authentication schema. An agent card adds a new layer for a new class of reader, the interactive agent. It provides structured capability declarations in JSON at a known address that robots.txt was never built to carry. The correct configuration uses both files serving their distinct purposes, not one replacing the other.
Do I need Salesforce to implement agent cards?
No. Agent cards follow the open A2A standard governed by the Linux Foundation and supported by more than 150 organizations. Any stack that can serve a JSON file at /.well-known/agent-card.json and expose MCP endpoints can implement a complete agent card. Salesforce Headless 360 is one implementation of this standard, built on top of the Salesforce Customer 360 ecosystem with per-conversation pricing. Independent implementations serve the same standard without the platform dependency or the CRM-first architecture requirement.
What happens if my agent card declares capabilities my MCP server does not fulfill?
A2A-compliant agents validate capability declarations against the MCP server during the discovery handshake. If a declared capability resolves to a non-functional or missing endpoint, the agent marks the capability as unavailable and may deprioritize the site in future discovery passes. Keeping the agent card and MCP server in sync is an operational requirement, not a one-time setup task. This is why autonomous maintenance matters, because a manually managed agent card drifts out of sync as endpoints change, tools are added, and authentication schemes rotate.
How does llms.txt differ from a sitemap.xml?
A sitemap.xml tells search engine crawlers which URLs exist on a site and when they were last modified. It is designed for traditional indexing pipelines. An llms.txt file tells LLM crawlers how to read and understand the site’s content, what the site is about, and how to structure citations. It operates at the semantic layer rather than the URL layer. Both files serve distinct audiences and should be published together. Omitting llms.txt leaves LLM crawlers without the structured context they need to produce accurate citations, even if every URL is correctly listed in the sitemap.
Conclusion
Agent cards now sit at the core of agentic technical SEO. The A2A standard is in production, the agentic AI market is growing rapidly, and the agents doing discovery, research, and booking on users’ behalf already read the web for structured capability declarations. A site without a complete agent card, a live MCP server, published llms.txt files, and Markdown serving for agent crawlers is undocumented to the systems that now decide which brands get cited and which get ignored.
AI Growth Agent is the only production-ready engine that ships the complete stack. That stack includes agent cards, Blog MCP, agentic technical SEO, living content, and zero-headcount headless marketing, all provisioned on the client’s own domain with the same rapid deployment described earlier. No Salesforce license. No agency dependency. No per-conversation billing. The first article goes live within a week, content indexes in as little as ten days, and the engine self-heals every card, every file, and every piece of content as the standards evolve.
