Clarify pydantic-ai tool chaining and voice-agent flows

[fbraza]

Summary

• document how pydantic-ai agents pass tool outputs via conversation history
• describe recommended pattern for chaining slack.user.resolve then slack.private.message
• capture voice-agent requirements for natural-language recipient selection

Acceptance Criteria

• README or docs section outlines tool chaining behavior and prompts
• guidance explains why resolver is a tool rather than injected object
• voice UX considerations (pronunciation, ambiguous matches) are recorded for future work

Notes

• reference findings from https://ai.pydantic.dev/llms-full.txt
• include examples of LLM transcript showing resolver → message flow
• mention limitations (no automatic tool chaining, LLM must orchestrate calls)

Example

### Example run
1. Agent calls `slack.user.resolve` with `{"names": ["jane"]}`
2. Tool returns `[{"requested": "jane", "user_id": "U123", "confidence": 0.93}]`
3. Agent calls `slack.private.message` with `{"users": ["U123"], "text": "Meeting at 3pm"}`
4. Delivery report confirms success

> Note: PydanticAI stores each tool response in the dialogue history, so the LLM can reuse fields in follow-up tool calls.