Architecture

​

Overview

The StackOne A2A server exposes every linked StackOne account as an Agent2Agent (A2A) agent at https://a2a.stackone.com. It is built on Google’s Agent Development Kit (ADK): an ADK Runner drives a Gemini model, the StackOne ADK plugin supplies the account’s tools, and the StackOne SDK runs those tools against the StackOne Actions API. The protocol itself is implemented with Google’s A2A Python SDK.

​

Request lifecycle

When a client sends a message to your StackOne agent:

1. Your client or agent sends an A2A message to the StackOne A2A server. For message/stream, responses stream back over Server-Sent Events.
2. The server works out which StackOne actions are needed and runs them against the accounts your credentials cover.
3. The result is returned in the response, or streamed back as it is produced.

The server optimizes how it discovers and runs the right actions to keep responses fast, and improves over time.

​

Components

​

A2A protocol layer

Google’s A2A Python SDK implements the A2A 0.3 surface: message/send, message/stream, the tasks/* methods, and the public and authenticated agent cards. Conversation and task state is persisted, so long-running tasks survive across requests.

​

Google ADK Runner

The Agent Development Kit Runner executes an agent on a Gemini model for each request, managing the model loop, the tool calls, and streaming. ADK is the foundation that lets StackOne agents interoperate with the wider Google agent ecosystem.

​

StackOne ADK plugin

The StackOne ADK plugin (stackone-adk) exposes the actions available to your account, across the 200+ providers in StackOne’s integration catalog, to the model as native ADK tools. The hosted server runs it in search-and-execute mode, registering just tool_search and tool_execute so the model discovers and runs actions at runtime rather than loading every tool up front. It is the same plugin StackOne provides to customers who build their own ADK agents, so the hosted agent and your own agents share one tool layer.

​

Search and execute

The plugin runs in a search-and-execute mode. Rather than listing every action up front, it gives the model two tools: one to search the action catalog and one to execute a chosen action. The model discovers tools on demand, so the prompt stays a constant size and accuracy holds as you link more connectors to your account. For how the search itself works, see Tool Search 101.

​

Why Google ADK

The server is built in Python on Google’s ADK so that StackOne A2A agents sit inside the Google agent ecosystem and interoperate cleanly with ADK-built agents and Google’s agent platforms. It uses StackOne’s own ADK plugin and SDK rather than bespoke integration code. Google actively maintains both ADK and the A2A Python SDK, with broader feature coverage than the TypeScript equivalents, so the server tracks the protocol as it evolves. When ADK adopts v1 of the A2A protocol, moving the StackOne server to it will be straightforward.

​

Next steps