Add LaunchDarkly AI Configs integration sample for Strands

[sattensil]

Summary

Adds a new sample under python/03-integrate/runtime-control/launchdarkly that drives a Strands agent entirely from LaunchDarkly AI Configs: model, instructions, parameters, tool list, and the multi-agent graph topology all live in LaunchDarkly.

The notebook is self-contained — it creates the project, AI Configs, variations, governed tool, default targeting, and an Agent Graph via the LaunchDarkly API on first run; re-running is idempotent.

Highlights:

• create_strands_model maps an AIAgentConfig to the right Strands model class (OpenAIModel / AnthropicModel / BedrockModel) by reading config.provider.name with a Bedrock model-id fallback — no per-provider branching in the sample.
• Three triage variations (OpenAI gpt-5, Anthropic Claude Sonnet 4.6, Bedrock Claude Sonnet 4.6) demonstrate provider swap via the LaunchDarkly UI.
• The agent's tool list comes from config.model.parameters['tools'], resolved at runtime against a local TOOL_REGISTRY. Detaching a tool in LaunchDarkly takes effect on the next invocation with no code change.
• Per-config tracking via tracker.track_metrics_of_async + an LDAIMetrics extractor fires duration / success / tokens atomically for async Strands invocations; track_tool_call fires from the @tool body.
• execute_graph walks node.get_edges() at runtime, builds a Strands Agent generically per node from node.get_config(), and routes between agents based on edge.handoff.route values the LLM emits. Graph-level handoff success/failure + the final path are tracked via graph.create_tracker().
• Routing protocol (JSON envelope, when to hand off) lives in each AI Config's instructions in LaunchDarkly; the builder only injects the runtime route list from edges. Adding a node + edge in the LaunchDarkly UI changes which agents fire on the next run.
• Section 9 demos both branches: a status-only query terminates at the triage agent; an analysis query hands off to a strands-specialist-agent.

Test plan

• pip install -r python/03-integrate/runtime-control/launchdarkly/requirements.txt
• Create a .env with LAUNCHDARKLY_API_TOKEN, OPENAI_API_KEY, ANTHROPIC_API_KEY (per README)
• Run the notebook top to bottom; verify section 1 creates the project, sections 2-4 create the variations + governed tool + default targeting, section 7 runs three multi-turn root-agent turns, section 8 creates the specialist + agent graph, section 9 runs both demo queries
• Confirm the Agent Graph view in LaunchDarkly shows non-zero invocations on both nodes after section 9
• Swap the default variation on strands-agent between gpt-5-agent, claude-sonnet-agent, bedrock-claude-agent in the LaunchDarkly UI and re-run section 9 — same code, different provider