Five tools with complete input schemas and descriptions. Naming is clear and action-oriented (search, deep_search, navigate, screenshot, read_page). Descriptions are detailed (100-250 chars) and include usage guidance. Parameters have types and descriptions. However, output schemas are not formally documented, responses are free-text strings rather than structured objects. Error handling is present but generic. No tool annotations (readOnlyHint, idempotentHint). Instructions are comprehensive but belong in system prompt, not tool definitions.

• Convert all tool responses from free-text strings to structured JSON objects. E.g., search() should return {results: [{title, url, snippet, published}], total_count, unresponsive_engines} instead of formatted text.
• Add tool annotations: mark search, deep_search, navigate, screenshot, read_page as readOnlyHint=true and idempotentHint=true so LLMs know they are safe to retry.
• Enhance error messages with recovery guidance. Instead of 'ERROR: search failed: ConnectionError', return 'Search failed: network timeout. Try again in 5s or broaden your query (remove time_range filter).'
• Move routing instructions and call budget rules from tool.instructions into a separate system prompt or agent configuration. Keep tool descriptions focused on WHAT each tool does, not HOW to orchestrate them.
• Add formal JSON Schema enum constraints for 'mode' (read_page), 'format' (navigate), 'wait_until' (navigate, read_page), and 'categories' (search, deep_search) to prevent invalid LLM inputs.