AI Agent with Tool Use — ReAct Pattern

## Task
Build an AI agent using the ReAct (Reasoning + Acting) pattern with tool/function calling.

## Requirements
- Language: TypeScript
- LLM: Claude or GPT-4 via API
- Pattern: ReAct loop (Thought → Action → Observation → repeat)

## Specifications
```
Tools available to the agent:
- search(query: string) → search results
- calculator(expression: string) → number
- readFile(path: string) → file contents
- writeFile(path: string, content: string) → success/failure

Agent loop:
1. Receive user query
2. Think: what do I need to do?
3. Act: call a tool
4. Observe: process the result
5. Repeat until answer is found or max iterations reached
```

## Implementation Notes
1. Use structured output for tool selection (JSON schema)
2. Implement retry logic for failed tool calls (max 3 retries)
3. Add conversation memory (last 10 turns)
4. Include a "give up" condition (max 5 iterations or explicit "I don't know")
5. Log each reasoning step for debugging
6. Handle tool errors gracefully — agent should recover, not crash

## Testing
- Test with a query requiring 2+ tool calls
- Test with an unanswerable query (agent should say it can't help)
- Test with a tool that returns an error