Registry probe of Pydantic AI's discovery endpoints

Pydantic AI · 2026-10-11 · AgentDossier refresh · harness probe-1.0

Scenario. Fetch and validate the ARD manifest, A2A agent card and MCP server card at the publisher's well-known paths

Observed. A2A verified (auth oauth2; unsigned); ARD verified; MCP invalid

Transcript

# Probe transcript — Pydantic AI

Agent `res_36273987eedd5e90` · publisher domain `pydantic.dev` · 2026-10-11 · harness probe-1.0

## Agent2Agent (A2A)

- requested `https://pydantic.dev/.well-known/agent-card.json`
- result: **verified**
- declared version: 0.3.0 (card 1.0.0)
- auth schemes: oauth2
- signed: no
- skills advertised: Query telemetry data, Manage dashboards, Manage alerts, Manage feature flags, Manage notification channels, Execute Python in a sandbox

## Agentic Resource Discovery (ARD)

- requested `https://pydantic.dev/.well-known/ard.json`
- result: **verified**
- manifest entries: 4

## Model Context Protocol (MCP)

- requested `https://pydantic.dev/.well-known/mcp.json`
- result: **invalid**

---
Observation only: the registry fetched public well-known paths and validated the documents against the protocol specifications. No credentials, no message exchange beyond discovery, nothing written.

Produced automatically on every refresh; a new bundle is written only when the observation changes

A trace describes one version on one date. It is evidence of behaviour, never a certification. Dispute it through the agent's dossier; corrections are appended, never overwritten.