Graph Orchestration: Weaver, Runner, and Evaluator

Arachne's core execution loop is driven by the interaction between the Weaver (planning), the Runner (execution), and the Evaluator (verification).

The Graph Weaver

The Weaver is a DSPy module responsible for transforming a high-level goal into a Directed Acyclic Graph (DAG) of discrete tasks.

Key Responsibilities

• Goal Analysis: Deconstructs complex goals into manageable nodes.
• Dependency Map: Determines the optimal order of execution.
• Tool Assignment: Maps required tools and MCP servers to specific nodes.
• Fail Context Incorporation: When re-weaving, it incorporates failure context to avoid previous pitfalls.

The Graph Runner

The Runner executes the DAG in "waves." A wave consists of all nodes that have their dependencies satisfied and can run in parallel.

Execution Process

1. Wave Identification: Find nodes with met dependencies.
2. Parallel Execution: Execute each node in the wave simultaneously.
3. Context Propagation: Pass the output of predecessor nodes as input to successor nodes.
4. Checkpointing: Save the state of each wave to disk immediately upon completion.

The Triangulated Evaluator

Once the Runner completes the graph, the Evaluator performs a multi-level check on the final output.

• Level 0 (Rules): Instant verification against numeric or hard constraints (e.g., "cost < $1.00").
• Level 1 (Semantic): A DSPy module scores the result's confidence against the success criteria (0.0 to 1.0).
• Level 2 (HITL): If confidence is below the threshold, the system flags the result for human-in-the-loop review.