Loading ToolBench
An AI agent framework with tool retrieval, delegation, and context management capabilities for orchestrating multi-agent workflows
IronHive Agent exposes only 2 tools with minimal schema documentation. The 'delegated_agent_tool' has a basic input schema with 2 parameters, but descriptions are present. Both tools have descriptions, but they are generic and lack actionable guidance on when/why to use them. No output schemas are documented. No error handling guidance. No parameter constraints (enums, ranges). The codebase shows sophisticated embedding-based tool retrieval (EmbeddingToolRetriever.cs) but this is internal infrastructure, not exposed as a tool. Tool definitions appear minimally specified for an MCP server.
Consult a stronger reviewer model about the task so far. Takes no input: it reads the whole conversation, including your tool calls and their results. Call it before committing to an approach on a non-trivial task, when you are stuck or results do not fit, and before declaring the task done. Weigh its advice seriously.
Delegates a sub-task to a named Ironbees agent. Each call is an isolated run of that agent with its own tools, prompt and model. The final answer comes back as the tool result.
advisor tool has NO input schema. Tool description states 'Takes no input' but this is not formally declared in a schema structure.
delegated_agent_tool input schema lacks type definitions for 'task' and 'context' parameters. Schema shows only descriptions, no 'type' field visible.
No output schemas documented for either tool. LLMs cannot plan downstream calls or extract required fields without knowing response structure.
Tool descriptions lack actionable guidance. 'Consult a stronger reviewer model' and 'Delegates a sub-task' are vague. No guidance on error recovery, prerequisites, or when to prefer one tool over another.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
No error handling guidance. If delegated_agent_tool fails, LLM has no recovery path. No categorization of errors as retryable, user-fixable, or fatal.