IBM Research Uses AI “Agent Factories” to Design Quantum Error-Correction Decoder Hardware
In an Oct. 13 paper, IBM Research reports that a team of AI agents optimized hardware for a quantum error-correction decoder. Open-weight models served on-premises reached a 52× estimated speedup for about $26 in inference; a proprietary setup reached 204× for about $299.
By 1EyeAI Desk ·
IBM Research reported on Oct. 13, 2026, in a blog post and preprint, on using an AI agent factory to optimize classical hardware that decodes quantum error-correction codes. The target was a High-Level Synthesis (HLS) C++ implementation of the Relay-BP decoder for the [[144,12,12]] bivariate bicycle code. Authors include Abhishek Bhandwaldar, Mihir Choudhury and IBM Research chief scientist Ruchir Puri.
A coordinator agent assigned optimization ideas to four parallel worker agents, with correctness checks and HLS results guiding each round. Each configuration got a budget of 100 hypotheses. Changes included reorganizing message storage and replacing sequential reductions with tree structures to expose parallelism.
IBM’s reported results, based on HLS estimates rather than a fabricated chip: an all-proprietary configuration reached a 204× latency speedup for about $299 in inference; a hybrid reached 126× at 2.8× lower cost; an all-open-weight setup served on-premises with the open-source llm-d stack reached 52× for about $26, 11.5× cheaper. Doubling the open budget to 200 hypotheses raised the speedup only to 60×. IBM reports a ~98% prefix-cache hit rate and blended serving cost of about $0.11 per million tokens. Quantum Computing Report identifies the models as Claude Opus 5 and GLM-5.3.
IBM notes each configuration was evaluated in a single search run, so the results “characterize the observed searches rather than establish a universal model ranking,” and costs cover inference only, not the full hardware workflow.
What this means for you: This is research, not a shipping chip. If you run agent-heavy engineering work, the takeaway is that serving infrastructure can matter as much as model choice for cost.