a2a cloud
sign inStart deploying
honest comparison

You can build this yourself.
Most teams shouldn't.

Each tool below is good at what it's good at. None of them are an agent runtime. That's the gap.

stack vs runtime

What each tool leaves you to build.

LangGraph / LangChain

orchestration library

good at: Graph-shaped agent logic. Great for in-process flows.

you still build
  • No service identity. Agents aren't deployable as standalone services.
  • No scoped grant primitive between graph nodes.
  • No AgentCard, no MCP surface, no /tasks protocol surface.
  • Receipts are your problem.

CrewAI / AutoGen / swarm libs

multi-agent framework

good at: Role-shaped collaboration loops. Fast to prototype.

you still build
  • Same in-process trap — agents are objects, not services.
  • Tools share the host process's filesystem and creds.
  • No discovery standard. No marketplace path.
  • Audit is whatever you logged.

DIY FastAPI + auth

hand-built service

good at: Full control. Familiar shape.

you still build
  • You will rebuild AgentCard, MCP, file grants, receipts, replay.
  • Tenanted auth, key rotation, sandboxing are not freebies.
  • Every team will rebuild it differently.
  • Switching from prototype to governed prod is a rewrite.

Cloud Run / Fargate / Lambda

managed compute

good at: Solid container or function runtime. Good autoscale.

you still build
  • Compute primitives. No agent primitives.
  • No standard for skills, schemas, AgentCards, pricing.
  • Workspace + grant + handoff are above this layer.
  • MCP/A2A interop is your integration job.

Kubernetes + service mesh

platform plumbing

good at: Tenant isolation, mTLS, policy — at the network layer.

you still build
  • Network identity ≠ agent identity. No AgentCard.
  • No file grants, no receipts, no replay surface.
  • Months of platform work before the first agent ships.
  • You're still inventing a deploy DX on top.

Internal home-grown

the inevitable Notion doc

good at: Shaped to your stack. Maybe.

you still build
  • Three teams will build three of them.
  • None will have receipts or grant primitives.
  • Hiring will require explaining a snowflake.
  • External agents can't speak your protocol.
what a2a cloud is

A runtime that ships with agent primitives, not just compute.

The differences below aren't individual features. Each is a primitive the runtime owns — so the same building blocks work across teams, agents, and external partners.

service identity per agentyes
AgentCard + skill schemasyes
scoped grants (HMAC, audience-bound, time-limited)yes
sandboxed code executionyes
workspace files with grant boundaryyes
a2a deploy → managed runtimeyes
A2A protocol surface (tasks, messages, artifacts, SSE)yes
MCP bridge for every skillyes
receipts: who, what, how, proofyes
replay + eval hooksyes
per-call pricing ledgeryes
private registry + public discoveryyes
inevitability

Standardize the runtime before every team builds a different one.

The same dynamic that produced Kubernetes, Docker, OAuth, and service meshes is happening for agents. Build it once, on a standard. Or build it three times, by accident, in three teams.

Install the SDKView architecturePlan a pilot