Loading ToolBench
A file-first agent operating system. Build AI agents by editing markdown files, not writing code.
Agent Harness exposes three tools with severe definition gaps. agent_tools_delegation has a minimal description and vague semantics around error handling. http_tool_operations lacks parameter descriptions entirely, 'body' and 'query' are undescribed strings with no format guidance, and the tool name itself is generic ('http_tool_operations') rather than verb-noun. programmatic_tools has an empty schema (no properties documented) and no description. None of the tools document output schemas, error recovery paths, or constraints. Tool names do not follow verb_noun convention. No tool includes parameter type validation hints, enum constraints, or actionable error messages. This server is not production-ready for agent use.
Delegate to a sub-agent defined in agents/*.md. Tool name equals the agent's frontmatter id. Input is {prompt: string}. Returns the subagent's final text or an error object on failure.
Execute HTTP tool operations defined in tools/*.md. Each operation becomes a separate tool entry with name format {toolId}_{operationName}. Supports GET, POST, PUT, DELETE, PATCH with URL parameter resolution, query parameters, request body, and authentication headers.
Execute programmatic tools defined via ProgrammaticTool interface. Tools have name, description, inputSchema (Zod type), and async execute function.
http_tool_operations has zero parameter descriptions. 'body' and 'query' are bare strings with no format, encoding, or validation hints. LLMs cannot infer whether 'body' expects JSON, form-encoded, or raw text.
programmatic_tools has an empty input schema (no properties) and no description. This tool is effectively undocumented and unusable by agents.
Tool names do not follow verb_noun convention. 'http_tool_operations' and 'programmatic_tools' are generic nouns, not action verbs. 'agent_tools_delegation' is awkward. Rename to 'delegate_to_agent', 'execute_http_request', 'execute_programmatic_tool'.
No output schemas documented for any tool. Agents cannot plan downstream calls or extract required fields (e.g., does http_tool_operations return status_code, headers, body separately or as a single string?).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 31 | 2026-07-28+ | v2 |
Error handling is minimal. agent_tools_delegation returns {error: message} on failure, but does not guide recovery (e.g., 'Try calling search_agents() first'). http_tool_operations and programmatic_tools have no documented error paths.