Tools and Multi-Protocol Architecture
Tools are how the agent does things beyond generating text — fetching a page, running a calculation, charting data, calling a remote service. The platform is multi-protocol: a single agent loop can call tools that live in very different places and authenticate in very different ways, all behind one uniform tool interface.
Multi-protocol support
Section titled “Multi-protocol support”The agent can reach a tool through any of four protocols. Each trades locality for reach: a direct Python function is the simplest and fastest; an A2A agent is a fully independent service you delegate to.
| Protocol | Where the tool runs | Auth | Status |
|---|---|---|---|
| Direct call | In-process Python function in the agent (agents/main_agent/tools/) |
None — same process | Available |
| AWS SDK | In-process, but calling an AWS service via boto3 |
IAM (the runtime’s task role) | Available |
| MCP + SigV4 | A Lambda behind the AgentCore Gateway, exposed as MCP | AWS SigV4 | Available |
| A2A | A separate agent in its own AgentCore Runtime | AgentCore auth | Client-only today |
A few notes on the edges:
- Direct and AWS SDK tools are ordinary
@tool-decorated functions. The difference is only what they touch — local computation versus an AWS API. - MCP + SigV4 tools are discovered dynamically from the Gateway, so the tool list grows without a code change. See Gateway MCP Targets for registering one.
- A2A is client-only right now: the platform can call out to remote agents, but does not yet expose itself as an A2A server.
An intentionally limited default set
Section titled “An intentionally limited default set”The platform ships with a deliberately small built-in tool set. This is a design choice, not an omission — the goal is a clean, dependency-light starting point you extend for your own deployment rather than a kitchen sink you have to prune.
What’s included out of the box:
| Tool | Protocol | What it does |
|---|---|---|
| Calculator | Direct (Strands built-in) | Evaluates mathematical expressions. |
| URL Fetcher | Direct | Fetches and extracts text from web pages, articles, and docs. |
| Charts & Graphs | Direct | Builds interactive bar, line, and pie charts from data. |
| Code Interpreter | Direct (sandboxed) | Runs Python in a sandbox to generate diagrams and visualizations. |
| Spreadsheet tools | Direct (sandboxed) | Lists and analyzes spreadsheet files from the knowledge base or attachments. |
The built-in registry is assembled in create_default_registry()
(agents/main_agent/tools/tool_registry.py); each tool’s display metadata lives
in the TOOL_CATALOG (tool_catalog.py).
Flexible by design
Section titled “Flexible by design”The small default set is paired with several extension points, so capability is something you add rather than something you’re stuck with:
- Add a code tool — drop a
@toolfunction intoagents/main_agent/tools/, register it in__init__.py, and it joins the registry. (Direct or AWS SDK.) - Add a remote tool without deploying agent code — register a Gateway MCP target; its tools are discovered at runtime and merged into the catalog.
- Delegate to another agent — point the agent at a remote A2A agent for whole workflows it can hand off.
- Control who sees what — tool visibility is governed by RBAC and per-agent
enabled_tools, filtered through theToolFilter. See RBAC and Permissions. - Curate from the admin console — the Tools admin page manages the catalog and Gateway targets without touching code.