Loading ToolBench
Pydantic-AI based Multi-Agent Framework with YAML-based Agents, Teams, Workflows & Extended ACP / AGUI integration. Exposes tools via MCP protocol for ACP agents.
Scoring was not performed
Tool name 'agent_cli' is a noun phrase, not verb_noun pattern. Does not convey actionable intent (e.g. 'execute_agent_command', 'manage_agent'). LLMs cannot infer what to do from 'agent_cli' alone.
Description is 93 chars but vague: 'Tool for executing internal agent management commands. Provides access to the agent's internal CLI for management operations like creating agents, managing tools, connecting nodes, etc.' Lists examples but does not enumerate valid commands, making it impossible for LLMs to determine what inputs are safe or valid.
Input parameter 'command' is a freeform string with no enum, no allowed values list, no regex pattern, and no constraint description. LLMs will hallucinate arbitrary command strings. Should enumerate valid commands or redirect to a separate discovery tool.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 34 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 11 | - | v1 |
Parameter 'ctx' (AgentContext) has no description. It is a complex object type with unknown fields. Completely opaque to an LLM, cannot determine what it contains, how to populate it, or whether it must be pre-initialized.
No output schema documented. The tool returns unknown structure, LLMs cannot plan downstream tool chains or extract required fields (e.g. agent_id, team_id, status).
Tool accepts arbitrary 'command' strings with leading slash optional. This is a code-injection risk if the command string is not validated server-side. No error handling or validation guidance provided to LLMs.
No error messages, recovery guidance, or error classification. If a command fails, the LLM receives no actionable feedback: is it retryable? Should I ask the user? Is it a syntax error or permission error?
Tool combines multiple responsibilities ('creating agents, managing tools, connecting nodes') into a single opaque tool. Violates single-responsibility principle. Should be split into separate, composable tools like 'create_agent', 'add_tool_to_agent', 'connect_agent_nodes'.