A2AConnect
Hand off work across any agent runtime — without sharing an SDK.
The SDK-agnostic interop layer for multi-agent fleets. LF A2A–compatible. Complements MCP. Python runtime. Site at a2aconnect.dev.
How it fits together
Peers exchange typed envelopes over a pluggable transport. Adapters wrap your existing brains. Policy and audit sit at the boundary. LF A2A for peers on the wire; MCP for tools.
- Discovery — hello / hello.ack, then capabilities.query → announce.
- Negotiation — offers list required capabilities; mismatches reject cleanly.
- Handoff — offer → accept → start → progress/events → result.
- Governance — policy allowlists and audit trails for production.
Agent brains
Your runtimes stay yours — Python, LangGraph-shaped graphs, CrewAI-shaped roles, OpenAI agents, or custom asyncio.
Runtime adapters
RuntimeAdapter → TaskHandler via from_async_fn. No hard SDK lock-in; wrap the brain you already run.
AgentPeer + governance
Handshake, capability negotiation, offer/accept, progress, events, result. Policy allowlists and MemoryAuditLog sit at the boundary.
Pluggable transport
Same envelopes over in-memory bus (dev), outbound buffer (backpressure), or HTTP JSON stub (edge).
Local interop profile
Typed JSON messages mapped conceptually to LF A2A task lifecycle. Complements MCP for tools — does not replace either standard.
LF A2A
Industry peer wire standard. A2AConnect maps profile ≈ Agent Card, task.* ≈ Task lifecycle — does not replace it.
MCP
Agent ↔ tools/data. Orthogonal vertical: use MCP for tools, A2AConnect for peer handoff.
Get started
Install the package, run the research-handoff quickstart, then the full cross-runtime demo (orchestrator → researcher → writer).
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest
python demos/support_desk.py # support agent→agent
python demos/incident_handoff.py # incident agent→agent
python demos/two_agents.py # cross-runtime brief
python demos/quickstart.pyimport asyncio, json
from a2aconnect import AgentPeer, MemoryAuditLog, create_bus
from a2aconnect.adapters import from_async_fn
async def research_run(input, ctx):
topic = (input or {}).get("topic", ctx.offer["goal"])
await ctx.report_progress(percent=40, message=f"Researching {topic}")
await ctx.emit_event(
name="source.found",
data={"title": f"{topic} overview"},
)
return {
"topic": topic,
"findings": [
f"{topic}: fleets need a shared handoff layer",
"LF A2A is the wire standard; A2AConnect is the runtime",
],
}
async def main():
bus = create_bus()
audit = MemoryAuditLog()
researcher = AgentPeer(
profile={
"id": "researcher",
"name": "Researcher",
"capabilities": [{"name": "research.web", "version": "1"}],
},
transport=bus.create_transport("researcher"),
handlers={"research.web": from_async_fn(research_run)},
)
orchestrator = AgentPeer(
profile={
"id": "orchestrator",
"name": "Orchestrator",
"capabilities": [],
},
transport=bus.create_transport("orchestrator"),
audit=audit,
)
researcher.start()
orchestrator.start()
await orchestrator.hello("researcher")
outcome = await orchestrator.offer_task(
to="researcher",
goal="Research agent interop",
input={"topic": "agent interop runtimes"},
required_capabilities=["research.web"],
)
print(json.dumps(outcome["result"]["output"], indent=2))
asyncio.run(main())
# full file: demos/quickstart.pyWalk the full two-agent path on the examples page.
Open-core, sellable fleet control
OSS forever for the runtime. Charge when fleets go multi-team — not for inventing a wire protocol.
OSS
PyPI a2aconnect, local transports, policy/audit hooks.
Gateway
Auth, routing, rate limits, multi-tenant registry — roadmap.
Enterprise
SSO/SCIM, VPC, retention, SLA — roadmap.