Add Brave Search web-search partner extension (aws-brave/)

[SaRedfiche]

Summary

Adds a partner extension under aws-brave/ that gives the workshop's agents live web search via the Brave Search API, following the framework's partner-extension pattern — an AgentCore Gateway Lambda target surfaced through a Strands A2A "Research Agent." The core aws-only/ track is untouched.

What's included

• code/web_search_lambda/ — stdlib-only Brave client + Lambda handler (anycompany_brave_web_search); reads the API key from Secrets Manager at runtime.
• code/research_agent/ — web_search @tool, MCP tool schema, and the A2A server (cloned from the base Order/Refund agent pattern).
• code/tests/ — 20 unit tests (Brave client, Lambda handler, tool-arg builder, schema) — no AWS or network needed.
• workshop/l3-orchestration/6_web_search_agent.ipynb — end-to-end deploy lab: key entry → Lambda → Gateway target → MCP smoke test → Research Agent deploy (CodeBuild) → Registry → direct A2A invoke → cleanup.
• README.md — architecture, a native AgentCore Web Search vs. Brave (partner) contrast, security, cost, and cleanup.

How it wires in

Research Agent (A2A, Registry-discoverable)
  └── @tool web_search → AgentCore Gateway (MCP, Cognito JWT)
        └── Lambda: anycompany_brave_web_search
              ├── Secrets Manager: brave/search-api-key
              └── Brave Web Search API

Routing through the Gateway means the tool inherits L4 PII masking and L5 tracing, consistent with the base agents.

Testing

• cd aws-brave/code && python3 -m venv .venv && .venv/bin/python -m pip install -r requirements-test.txt && .venv/bin/python -m pytest tests/ -v → 20 passed.
• Deployed and validated end-to-end in a SageMaker Studio environment (us-west-2): the raw MCP tools/call and the deployed Research Agent both return live Brave results with citations.

Security notes

• Brave API key lives only in Secrets Manager; never in code or the notebook (the key-entry cell has a single paste target + sanity check).
• Web results are treated as untrusted (prompt-injection guidance in the agent's system prompt; L4 Guardrail masking applies via the Gateway).
• Lambda execution role scoped to GetSecretValue on the Brave secret only.