Summary
a2a-sdk is the official Python SDK for the Agent2Agent (A2A) protocol, maintained by the a2aproject org (backed by Google and a Linux Foundation technical steering committee with AWS, Cisco, IBM Research, Microsoft, Salesforce, SAP, and ServiceNow participating). A2A reached a stable v1.0 spec in April 2026 and has production integrations in Azure AI Foundry, Copilot Studio, and Amazon Bedrock AgentCore Runtime. It defines an agent-to-agent execution surface distinct from MCP's tool-calling model: a client sends a task/message to a remote agent (message/send / message/stream JSON-RPC methods, exposed client-side via the a2a.client module) which triggers that agent's own reasoning/tool-use and returns a Task/Message result, with streaming (SSE) and polling (tasks/get, tasks/cancel) variants. Server-side, an agent implements this execution surface via AgentExecutor.execute(context, event_queue).
This repository has zero instrumentation for it — no py/src/braintrust/integrations/a2a* directory, no wrapper, no nox session, no cassette directory, and no entry in py/pyproject.toml's [tool.braintrust.matrix].
This is comparable in scope to other agent-run execution surfaces this repo already instruments (e.g. the tool-calling execution surface covered by the standalone MCP Python SDK gap already filed in #567), but for cross-agent delegation rather than local tool invocation.
Weekly downloads
Weekly downloads: 2,761,407 (as of 2026-07-23; https://pypistats.org/packages/a2a-sdk). Cross-checked against https://pepy.tech/projects/a2a-sdk, which shows 11,263,287 downloads over the trailing 30 days (~2.6M/week), consistent with the pypistats figure.
Braintrust docs status
not_found. Checked https://www.braintrust.dev/docs/guides/tracing and https://www.braintrust.dev/docs/guides/tracing/integrations — neither the provider list ("OpenAI, Anthropic, Gemini, AWS Bedrock, Azure, Mistral, Together, Groq, and many more") nor the agent-framework list ("OpenAI Agents SDK, Claude Agent SDK, Pi Coding Agent SDK, Deep Agents, LangGraph, CrewAI, AutoGen, AgentScope, Google ADK, LiveKit Agents, Mastra, Pydantic AI, Strands Agents SDK") mentions A2A or Agent2Agent.
What needs to be instrumented
- Client-side:
a2a.client message-send/stream execution (JSON-RPC message/send, message/stream) — the primary "call a remote agent" entrypoint, analogous to an agent-run span.
tasks/get / tasks/cancel polling for long-running task status.
- Server-side:
AgentExecutor.execute(context, event_queue) — the entrypoint an A2A-compliant agent implements to handle an incoming task; wrapping this would capture the receiving agent's execution as a span.
Upstream sources
Local repo files inspected
py/src/braintrust/integrations/ — no a2a directory
py/src/braintrust/wrappers/ — no A2A wrapper
py/pyproject.toml [tool.braintrust.matrix] — no a2a-sdk entry
py/noxfile.py — no test_a2a session
- Repo-wide case-insensitive grep for
a2a[-_]sdk under py/ — zero matches
Summary
a2a-sdkis the official Python SDK for the Agent2Agent (A2A) protocol, maintained by thea2aprojectorg (backed by Google and a Linux Foundation technical steering committee with AWS, Cisco, IBM Research, Microsoft, Salesforce, SAP, and ServiceNow participating). A2A reached a stable v1.0 spec in April 2026 and has production integrations in Azure AI Foundry, Copilot Studio, and Amazon Bedrock AgentCore Runtime. It defines an agent-to-agent execution surface distinct from MCP's tool-calling model: a client sends a task/message to a remote agent (message/send/message/streamJSON-RPC methods, exposed client-side via thea2a.clientmodule) which triggers that agent's own reasoning/tool-use and returns aTask/Messageresult, with streaming (SSE) and polling (tasks/get,tasks/cancel) variants. Server-side, an agent implements this execution surface viaAgentExecutor.execute(context, event_queue).This repository has zero instrumentation for it — no
py/src/braintrust/integrations/a2a*directory, no wrapper, no nox session, no cassette directory, and no entry inpy/pyproject.toml's[tool.braintrust.matrix].This is comparable in scope to other agent-run execution surfaces this repo already instruments (e.g. the tool-calling execution surface covered by the standalone MCP Python SDK gap already filed in #567), but for cross-agent delegation rather than local tool invocation.
Weekly downloads
Weekly downloads: 2,761,407 (as of 2026-07-23; https://pypistats.org/packages/a2a-sdk). Cross-checked against https://pepy.tech/projects/a2a-sdk, which shows 11,263,287 downloads over the trailing 30 days (~2.6M/week), consistent with the pypistats figure.
Braintrust docs status
not_found. Checked https://www.braintrust.dev/docs/guides/tracing and https://www.braintrust.dev/docs/guides/tracing/integrations — neither the provider list ("OpenAI, Anthropic, Gemini, AWS Bedrock, Azure, Mistral, Together, Groq, and many more") nor the agent-framework list ("OpenAI Agents SDK, Claude Agent SDK, Pi Coding Agent SDK, Deep Agents, LangGraph, CrewAI, AutoGen, AgentScope, Google ADK, LiveKit Agents, Mastra, Pydantic AI, Strands Agents SDK") mentions A2A or Agent2Agent.What needs to be instrumented
a2a.clientmessage-send/stream execution (JSON-RPCmessage/send,message/stream) — the primary "call a remote agent" entrypoint, analogous to an agent-run span.tasks/get/tasks/cancelpolling for long-running task status.AgentExecutor.execute(context, event_queue)— the entrypoint an A2A-compliant agent implements to handle an incoming task; wrapping this would capture the receiving agent's execution as a span.Upstream sources
Local repo files inspected
py/src/braintrust/integrations/— noa2adirectorypy/src/braintrust/wrappers/— no A2A wrapperpy/pyproject.toml[tool.braintrust.matrix]— noa2a-sdkentrypy/noxfile.py— notest_a2asessiona2a[-_]sdkunderpy/— zero matches