Agent Search has 9 tools with detailed descriptions (avg 250+ chars, well above 194-char baseline) and comprehensive parameter documentation. However, critical gaps exist: NO explicit input schemas visible in source code (only docstring-level descriptions), NO output schemas documented, NO error handling guidance, and NO tool annotations. Descriptions are verbose and helpful for humans but lack LLM-optimized structure. Tools are well-named (verb_noun pattern) and composition is sound, but schema formalization and error recovery patterns are missing.

• Add explicit JSON Schema definitions for all tool inputs. Use fastmcp's schema support or manually register via MCP protocol. Include type, description, enum constraints, and min/max for all parameters.
• Document output schemas for each tool. Specify return type (object/array), field names, types, and whether fields are always present or optional. E.g., web_search returns array of {title: string, url: string, snippet: string, rank_score: number}.
• Implement error classification: return {error: string, error_type: 'retryable'|'user_fixable'|'fatal', recovery_hint: string}. E.g., 'DeepSeek API unavailable. Retry in 30s or use web_search without summarization.'
• Add tool annotations via MCP protocol: mark all tools with readOnlyHint=true since they do not modify state. This enables agents to reason about safety.
• Expand web_map description to 100-150 chars explaining when to use it (site discovery before deep crawl), expected output (link tree), and performance (fast, single-level). Distinguish from web_crawl.