AgentProp：用图论和度量维度实现多智能体工作流的精准控制

AgentProp is an open-source project (Apache 2.0 license, v0.1.0a3) that models multi-agent AI workflows as directed weighted graphs. It leverages metric dimension theory and random zero-forcing propagation algorithms to achieve fault location, verifier optimization placement, and runtime control. In practical tests, it reduces token consumption by 33.8% and cost by 41% while maintaining task success rates.

Traditional multi-agent systems often use simple chain or tree structures, lacking systematic topology analysis. When an agent produces an error, it propagates downstream, but a verifier's failure may have multiple upstream root causes—leading to ambiguous fault localization. AgentProp solves this by introducing graph theory's metric dimension concept to provide mathematically provable fault diagnosis guarantees.

• Metric Dimension & Resolving Set: Verifiers act as 'landmarks'; a resolving set ensures each node has a unique distance vector to verifiers, enabling precise fault location (even with one verifier failure via fault-tolerant metric dimension).
• AgentGraph: Directed weighted graph with nodes (agents, tools, verifiers, etc.) and edges (weighted by information cost or failure probability), supporting serialization and visualization.
• Key Components: Propagation models (independent cascade, linear threshold, random zero-forcing, etc.), RZF centrality (for large graphs), runtime controller (adaptive strategies like retry/stop), and quality cascade model (correctness/compression propagation).

• Benchmark: Terminal-Bench 2.1 tests show 33.8% lower token use, 41% cost reduction, and 14.8% faster runtime (success rate maintained).
• Optimizations: Memoized distance calculations, lazy CELF seed selection, auto strategy choice based on graph size, etc.
• Integration: Supports LangGraph, AutoGen, CrewAI, etc. Quick usage via CLI (analyze/optimize workflows), Python API, or FastMCP server.

• Current Limitations: Alpha-stage software (early benchmark evidence, requires basic graph theory knowledge).
• Future Plans: Explore GNN-based propagation models, integrate reinforcement learning into controllers, expand validation across more benchmarks, and enhance visualization tools.

• Applicable Scenarios: Complex multi-agent systems, cost-sensitive applications, high-reliability systems, and audit-required workflows.
• Not Applicable: Simple linear workflows, prototype development, or resource-constrained environments.
• Industry Shift: Promotes moving from 'prompt engineering' to 'graph engineering'—emphasizing structural importance, theoretical guarantees, observability-first design, and cost-quality balance.