Agentica exposes 8 tools with significant quality gaps. Tool names follow verb_noun convention (search_, add_, read_, write_, apply_, execute, task, delegate), which is positive. However, descriptions are minimal (10-50 chars), parameter descriptions are sparse or missing, and input schemas lack proper type definitions and constraints. Most critically, dangerous tools like execute (shell command) and write_file lack any error handling guidance, confirmation patterns, or permission checks. No output schemas are documented. The server appears to be a custom HTTP gateway without MCP-native protocol support, tool definitions are inferred from code rather than explicitly registered via MCP mechanisms.

• Add confirmation/dry-run support to execute, write_file, and apply_patch. Example: execute() should accept a 'dry_run' boolean parameter and return what would be executed before actually running it.
• Expand all parameter descriptions to 50-150 characters with format/constraint guidance. E.g., 'model' → 'Model profile name (e.g. deepseek, openai). Must match a profile in ~/.agentica/config.yaml.'
• Document output schemas for all tools. E.g., search_knowledge_base should return {results: [{id, content, score}], total_count, next_cursor}.
• Add input validation and error recovery guidance. E.g., execute() should return 'Command failed with exit code 1: <stderr>. Retryable: false. Suggestion: check command syntax.' instead of raw error.
• Implement MCP protocol compliance: register tools via MCP's ListTools, expose input/output schemas in JSON Schema format, and support per-request _meta with logLevel.