Loading ToolBench
Parses workflow processes to automatically create specific sub-agent instructions and scaffold agent roles based on workflow descriptions
This server exposes a single tool with significant quality gaps. The tool name 'deconstruct_workflow' is appropriately verb-based, but the implementation reveals critical issues: (1) The input schema is minimal, only a single string parameter with a basic description. (2) No output schema is documented despite the tool returning a complex nested structure with 'steps_detected', 'scaffolded_agents', and per-agent fields like 'role', 'system_prompt', 'input_spec'. (3) The tool logic is entirely pattern-matching on keywords ('analyze', 'check', 'email', 'send', 'notify', 'order', 'buy', 'purchase') without any mechanism to handle workflows that don't match these hardcoded patterns, it silently falls back to a generic agent. (4) No error handling, validation, or recovery guidance. (5) The description lacks context on when to use this tool, what formats it accepts, or what the agent scaffolding output is meant for. (6) No pagination, limits, or constraint documentation. The server is NOT exposed as an MCP server, it is a plain Python class with no MCP protocol binding visible in the source. This is a local SDK, not a remotely-accessible agent tool.
Parses workflow description to automatically scaffold agent roles and detect workflow phases
No output schema documented. Tool returns a dict with nested structure (steps_detected list, scaffolded_agents list with role/system_prompt/input_spec per agent), but LLMs have no formal specification of what to expect. This forces LLMs to infer structure from one example call.
Tool description is 156 characters but lacks essential context. It does not explain WHEN to call this tool vs alternatives, WHAT the output scaffolding is used for, or HOW to use the returned agent definitions. Missing dependency hints and usage examples.
Logic is entirely keyword-matching with no fallback error handling. If a workflow description contains none of the hardcoded keywords ('analyze', 'check', 'email', 'send', 'notify', 'order', 'buy', 'purchase'), the tool silently returns a generic 'General Task Helper' agent. No validation, no feedback to the LLM that the parsing may have failed or been incomplete.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 31 | 2026-07-28+ | v2 |
Input parameter has no constraints or format guidance. The 'workflow_description' accepts any string with no length limits, required format, or validation rules. LLMs may pass empty strings, or extremely long descriptions that break the keyword-matching logic.
No MCP protocol binding detected. Source code shows a plain Python class (WorkflowAgentScaffolderClient), not an MCP server implementation. No Transport mechanism, no request/response envelope, no tool registration with MCP schema. This is a local SDK, not a remotely-accessible MCP agent tool.