Mark Ren spent the better part of thirty years watching engineers wrestle with the unglamorous minutiae of chip design: verification suites that took days to debug, analog layouts that demanded endless hand-tuning, RTL code reviews that felt more archaeological than inventive. So when he left NVIDIA Research last year, where he'd led the ChipNeMo project adapting large language models for semiconductor workflows, he had a narrow thesis in mind. The real automation opportunity wasn't in the cloud or on some vendor's centralized platform. It was behind the firewall, running on the same multi-vendor EDA tools his prospective customers already owned.
On Tuesday, Ren's San Jose startup Agentrys disclosed it had raised $24.5 million across two rounds: a $19.1 million seed led by Etna Labs and an earlier $5.4 million pre-seed from MediaTek. According to Dealroom.co, the seed round ranks in the 95th percentile for its geography over the past four years, a signal of investor appetite for infrastructure that sits between generic AI and the specialized, capital-intensive world of electronic design automation.
Agentrys sells what it calls Agentrys Studio, an on-premises platform that lets semiconductor teams build and refine AI agents to handle verification, register-transfer logic, analog design and physical-design tasks. The software layers atop customers' existing tool stacks from Cadence, Synopsys, Siemens and others. Critically, it keeps proprietary chip data inside the customer's own network, a non-negotiable for fabless designers and foundries skittish about sending trade secrets to third-party servers.
The company claims it has partnered with several top-tier fabless semiconductor companies, a major global foundry, and a handful of well-funded chip startups, though it declined to name any. It also reported scoring above 90 percent accuracy on NVIDIA's public CVDP verification benchmark, hitting between 90 and 96 percent on the toughest verification categories and 95.8 percent overall on RTL tasks. In one company demonstration, Agentrys ran an end-to-end autonomous workflow that took a 32-bit CPU from specification to sign-off-clean GDS without human intervention. Industry publication SemiWiki described the output as a compact, embedded-class processor produced by a multi-agent workforce, though the timeline and reproducibility of such demonstrations in production environments remain open questions.
Ren's background gives Agentrys unusual credibility in a field where most AI promises still outpace AI delivery. Before NVIDIA, he spent years at IBM Research, where he won a Corporate Award for high-performance microprocessor design closure. He co-authored DREAMPlace, NVCell and VerilogEval, tools that have since found their way into commercial use across the industry. His co-founders bring overlapping pedigrees: Yun-Da Tsai, Duo Ding and Wuxi Li have logged time at NVIDIA, Meta, AMD, Samsung, Google and Siemens, according to the company.
The competitive landscape is getting crowded, and quickly. ChipAgents, which has raised $134 million and claims more than 120 deployments, positions itself as an agentic harness that sits atop existing EDA engines, a similar architectural bet. The three incumbent EDA giants have also moved. Cadence, Synopsys and Siemens EDA each unveiled agentic-AI initiatives recently: Cadence introduced ViraStack AI Super Agent, Synopsys rolled out tokenized consumption models for agent workflows, and Siemens announced agentic transformations for custom IC design. Whether these offerings represent genuine step-function improvements or rebranded feature sets remains a matter of debate among engineers who have seen automation cycles come and go.
"Building production-grade agents that reliably automate real engineering work is far from easy," Ren said in the funding announcement. "That's the problem Agentrys exists to solve." The phrasing is careful, perhaps deliberately so. Chip design has resisted full automation for decades, not for lack of ambition but because the work involves judgment calls, trade-offs and domain knowledge that don't compress neatly into algorithms.
Penny Deng, a general partner at Etna Labs, framed the opportunity differently. She sees chip design as one of the first high-value domains where recursive self-improvement—agents that learn from their own outputs and refine their own performance—might actually become practical rather than theoretical. Brian Hsu, managing director of MediaTek Innovation Fund, called Agentrys "a builder solution that enables semiconductor engineering teams to customize and improve AI agents for chipset design workflows."
Agentrys is now hiring across research engineering, model training, infrastructure and customer-facing roles in San Jose, Austin and Taiwan. The company introduced its Agentic Design Automation framework at a recent Design Automation Conference and is a member of NVIDIA Inception, the chipmaker's startup accelerator program. How fast it can convert demonstrations into repeatable, revenue-generating deployments will determine whether it joins the ranks of foundational EDA infrastructure or remains a well-funded experiment in a notoriously conservative industry.
