Stateful Swarms: 39x Cost Reduction

A new paradigm called 'Stateful Swarms' has been released as open source to address the high cost of AI agents and the difficulty of tracking and auditing them.

Unlike existing agents that repeatedly read documents and lose information in the process, Stateful Swarms uses a structured Blackboard to maintain persistent, auditable memory. Specialized agents perform tasks and store their results in centralized memory, allowing knowledge to be built and queried cheaply after reading a document only once.

Key Performance Metrics (Harvey AI Legal Agent Benchmark):

• Stateful Swarms: Achieved a 83.74% benchmark pass rate at a cost of $1.30 per task.
• Existing SOTA (Harvey): Recorded a 10.4% benchmark pass rate at a cost of $50.90 per task.

This approach outperformed Claude Code's Opus agent not only in the legal field but also in tasks such as analyzing Datadog's 10-K report. All code, experimental settings, data, and reasoning processes have been released under the MIT License, allowing anyone to verify and use them.