Agent-Worlds provides 16 tools with schemas and descriptions, but quality is inconsistent. Most tools have descriptions (many in Chinese), but parameter documentation is sparse. Schemas are present but lack depth, many parameters missing descriptions, type constraints are minimal, and output schemas are undocumented. Naming follows verb_noun convention (observe_map, edit_map, invoke_npc, create_npc, read_file, write_file, edit_text, delete_file, add_task, complete_task, goto_location, arrived_at, end_conversation, edit_memory_note, send_qq_notify, run_command), which is good. However, several tools lack clarity on when to use them vs. similar tools (e.g., edit_text vs. write_file), and error handling guidance is absent. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels assigned. Output schemas are not documented, agents cannot predict what fields to expect. Security concerns: run_command and delete_file are high-risk but lack confirmation patterns or dry-run support.

• Document output schemas for all 16 tools. Specify field names, types, and whether results are paginated. Example: 'observe_map returns {objects: [{id, name, description, x, y, width, height, color}], total_count}'. This enables agents to chain tools correctly.
• Add tool annotations to all tools. Mark read-only tools (observe_map, read_file) with readOnlyHint=true; destructive tools (delete_file, run_command) with destructiveHint=true; idempotent tools (write_file with mode=overwrite) with idempotentHint=true. This guides agent retry logic.
• Implement confirmation pattern for delete_file and run_command. Add a 'confirm' boolean parameter (default false) or a separate 'confirm_delete_file' tool that requires explicit approval before executing. Prevents accidental data loss.
• Expand parameter descriptions. For invoke_npc, explain what 'wait=false' means (async execution?). For create_npc, document 'llm_channel' and 'llm_model' (which LLM providers/models are supported?). For edit_text, clarify when to use 'find' vs 'insert_after' vs 'range'.
• Add error recovery hints to tool descriptions. E.g., read_file: 'If file not found, check the path or use run_command to list directory contents.' This guides agents toward self-correction without extra discovery calls.