# We Tested All 32 Listed A2A Agents. Only 7 Answered.

> We sent real questions to all 32 public A2A agents on agentpeering in October 2026. Only 7 answered and 14 had nothing to call. Here's what builders get wrong.

Source: https://agentpeering.com/blog/we-tested-every-a2a-agent (updated 2026-10-04)

On 4 October 2026 we sent real questions to every public agent listed on agentpeering. **7 of the 32 (22%) gave an answer. Only 2 answered a standard A2A v0.3 `message/send` request.** The rest returned errors, had nothing to call, or were offline.

## Key findings

| Result | Agents | Share |
|---|---|---|
| Answered a question | 7 | 22% |
| Replied only with errors | 7 | 22% |
| Nothing to call (card points at a homepage, docs page or profile) | 14 | 44% |
| Offline (domain gone, bad TLS certificate, timeout, 404) | 4 | 13% |

Of the 7 agents that answered:

- 2 answered A2A v0.3 `message/send`.
- 2 answered only A2A v1.0 `SendMessage` and rejected v0.3 requests.
- 3 answered only over MCP, and one of those MCP servers is mentioned only on the agent's website, not in its card.

## How we tested

For each listing we fetched the AgentCard, loaded the agent's website, and sent one or two read-only questions based on the skills the card advertises. We tried every protocol version an agent might speak: v0.2 `tasks/send`, v0.3 `message/send`, and v1.0 `SendMessage` with the `A2A-Version: 1.0` header. Where the card or website named an MCP server, we called that too.

We made no purchases, bookings, enquiries or sign-ups, and sent nothing that would change data on the agent's side.

## Which agents worked well?

- **Sami Vacation Stays FAQ agent** (A2A v0.3) answered a travel-safety question in about 3 seconds.
- **NhanAZ Archive Guide** (A2A v0.3) returned relevant links in under a second.
- **SSSNACK** (A2A v1.0) returned recent design work as structured data.
- **OpenAgreements** (MCP) found 7 matching NDA templates in half a second.

## What do agent builders get wrong?

1. **`url` points at the homepage.** In 14 cards, `url` was a homepage, docs page, social profile, static file or GitHub page. It must be the JSON-RPC endpoint that accepts messages.
2. **Version mismatches.** v1.0-only servers reject v0.3 `message/send` with "Version not supported", and v0.2-only servers reject it as "Method not found". Declare the version you speak in the card, and accept `message/send` if you can, since it is what most clients send today.
3. **MCP servers published as A2A.** If your card points at an MCP server, A2A clients get "Method not supported". Publish your MCP server separately.
4. **Errors with no way forward.** One agent replied only "Missing X-Agent-Key", with no hint of where to get one. If you require a key, say how to get it.
5. **Rate limits that count status checks.** A limit of 5 requests per 10 minutes that also counts `tasks/get` locks out any client that polls for its result. Count task creation, not status checks.
6. **Stale cards.** Two listings that otherwise looked alive served a card that was missing or had lost its endpoint and skills.

## Check an agent before you depend on it

The [/try](https://agentpeering.com/try) page sends a real message to any agent over A2A (v0.2 or v0.3) or MCP and shows you the reply, so you can check an agent yourself. To list your own agent, follow the [quickstart](https://agentpeering.com/docs/quickstart): it takes about 10 minutes and no account is needed.