SDK-first, MCP-native agent orchestration framework with visual DAG builder and LLM integration
ReasonFlow exposes 13 tools via FastAPI HTTP endpoints with basic Pydantic schemas. Tool names follow verb_noun convention (node_types, providers, validate, run, stream, result, list_workflows, get_workflow, save_workflow, delete_workflow, list_traces, get_trace, test_provider). Descriptions are present but brief (avg ~80 chars, below the 194-char baseline). Input schemas are visible and typed, but output schemas are undocumented, LLMs cannot predict response structure. Parameter descriptions are minimal or absent. Error handling is basic (HTTPException with status codes, no recovery guidance). No tool annotations (readOnlyHint, destructiveHint). Security: no secrets in params (good), but no permission gates or audit logging visible.
Delete a saved workflow definition by name
Retrieve a detailed execution trace from a pipeline run
Retrieve a saved workflow definition by name
List execution traces from recent pipeline runs (up to 50 most recent)
List all saved workflow definitions with their modification timestamps
Retrieve the catalog of available node types (LLM, Code, Decision) with their field definitions for the visual builder
List available LLM providers and their models, including Ollama and cloud providers
Output schemas undocumented. LLMs cannot predict response structure for any tool. E.g., run() returns {run_id}, stream() yields SSE events, result() returns {success, state, total_cost, tokens, error, trace}, but none of this is declared in tool definitions.
Parameter descriptions missing or minimal. 'nodes' and 'edges' in validate() and run() lack detail on structure, required fields, or constraints. 'name' in get_workflow() has no description. LLMs cannot infer what values are valid.
No tool annotations. delete_workflow is destructive but lacks destructiveHint. stream() and result() are read-only but lack readOnlyHint. Agents cannot infer safety/idempotency without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | 2026-07-28+ | v2 |
Retrieve the final result of a completed DAG pipeline execution
Execute a DAG pipeline with provided inputs and stream execution events back to the client
Save a new workflow definition or update an existing one
Stream execution events for a running DAG pipeline via Server-Sent Events
Test connectivity and functionality of a specified LLM model
Validate a DAG graph structure (nodes and edges) for correctness before execution
Error handling lacks recovery guidance. HTTPException(status_code=404, detail='Unknown run_id') tells LLM nothing about what to do next. No categorization of retryable vs fatal errors. No suggestions for alternative actions.
Destructive operations (delete_workflow) lack confirmation/dry-run pattern. No mechanism to prevent accidental deletion. Agents can irreversibly delete workflows without safeguards.