Irys introduces Stateful Swarms, an open-source paradigm for AI agents using structured blackboard memory to improve performance and reduce cost. On Harvey AI's Legal Agent Benchmark, it achieved an 83.74% criteria pass rate at $1.30 per task, compared to the state-of-the-art 10.4% at $50.90.
Hey Reddit. I'm Devansh, from Irys. Through our work, we've observed that Agents have 2 main issues: 1. They're very expensive to run. 2. They can be very hard to trace and audit (so you don't know how they come up with the answers they do). We're open sourcing a paradigm to solve these problems called "Stateful Swarms,". Simply put: instead of AI agents repeatedly rereading documents and losing information, Stateful Swarms use a structured blackboard to maintain persistent, auditable memory. Specialized agents perform specific tasks and store their results into this centralized, structured memory—meaning you pay once to read and understand your documents and then cheaply query and build upon that knowledge indefinitely. Using typing and implementing a degree of structiure allows us to maintain the blackboard in prod, ensuring that it doesn't grow unweildy (which tends to happen with current generation of memory solutions). Here's how it performed: * On Harvey AI’s Legal Agent Benchmark, we hit an 83.74% criteria pass rate and a 17.75% strict all-pass rate at just $1.30 per task. The current state of the art is Harvey’s published at 10.4% at $50.90 per task, so swarms are both better and cheaper. * We generalized beyond legal, analyzing Datadog's 10-K filings to produce a comprehensive investment memo, while Claude Code's Opus agents couldn't handle the context load and failed. Because we're committed to open science, we've open-sourced everything—the code, experimental setups, data, and full reasoning traces—under an MIT license. This lets you validate our claims directly, improve the approach, or adapt it for your own applications. We strongly believe the future is about AI systems that don't forget as they learn. If this resonates with you, come collaborate or build upon what we've started. Let's advance stateful, intelligent systems together. Whitepaper on the thesis here: [https://github.com/dl1683/ant-irys/blob/master/whitepaper.pdf](https://github.com/dl1683/ant-irys/blob/master/whitepaper.pdf) **Repo:** [https://github.com/dl1683/ant-irys](https://github.com/dl1683/ant-irys) A primer to the thesis here: [https://www.linkedin.com/pulse/stateful-swarms-make-ai-agents-cheaper-safer-better-devansh-devansh-8enxe](https://www.linkedin.com/pulse/stateful-swarms-make-ai-agents-cheaper-safer-better-devansh-devansh-8enxe)
Cursor's new agent swarm design uses planner and worker models to decompose tasks into a tree structure, achieving significant cost savings and quality improvements. In a test rebuilding SQLite from scratch in Rust, the new swarm reached 80% pass rate in four hours while the old swarm failed.
The author discusses the growing use of agent swarms/workflows for processing unstructured data at scale, noting that reliable execution drops significantly when deploying more than 30+ sub-agents in parallel, and teases a solution for combining intelligent decision-making with reliable task execution.
The author built a unified workspace tool for debugging multi-agent AI systems, tracking agent state and loops with automated detection of infinite tool loops and context inflation.