Microsoft Agent Framework has released channel packages that expose one agent or workflow through Responses, Telegram, A2A and MCP; carrying a session between them depends on identity mapping owned by the application. Anthropic has meanwhile moved organisation administration into the ant command-line tool and seven language SDKs, while usage and cost reports remain direct-HTTP only.
Microsoft Agent Framework's new Python packages expose an agent or workflow to OpenAI Responses clients, Telegram, other agents using A2A, and MCP clients. A shared hosting layer holds the target and session state, while each channel translates input and output at its own protocol boundary. A developer can therefore connect the same agent to one or several channels without rebuilding it as separate applications maintained through separate SDKs.[1]
Shared history does not appear automatically. The application maps an identity from a Responses call or Telegram to a canonical session ID; if two channels resolve to the same ID, they read and update the same AgentSession, while different IDs keep their histories apart. Identity linking, authorisation and concurrency control remain the application's responsibility. Telegram commands and media handling, the A2A task lifecycle, and registration with an MCP server also stay in application code beyond the channel adapters.[1]
On the same day, Anthropic added its Admin API to the ant command-line tool and the Python, TypeScript, C#, Go, Java, PHP and Ruby SDKs. These surfaces cover organisation information, members, invites, workspaces, API keys, rate limits, service accounts and customer-managed encryption keys. Usage and cost reports, along with some Claude Enterprise management and analytics endpoints, remain accessible through direct HTTP calls. Compliance API endpoints for Cowork and Claude Code session transcripts also moved out of beta.[2]
