Agentrys raises $24.5M: agents that own a whole workflow

The interesting thing about the Agentrys round is not the dollar figure. It's the category the company is claiming: not a smarter tool for one task, but a system that owns an entire engineering workflow end to end. That distinction is the one that decides whether vertical AI agents save you money or just move the work around.

What actually happened

Agentrys announced $24.5 million in funding on August 27 — an oversubscribed $19.1 million seed led by Etna Labs, on top of a $5.4 million pre-seed led by MediaTek, its first strategic backer. Semiconductor Digest reported the money goes to hiring, agent-native tooling, and customer work across verification and physical design.

The company calls the category agentic design automation: where traditional EDA tools automate individual tasks, Agentrys is building systems that run, learn from, and improve whole workflows. The proof point it offers is an autonomous multi-agent run that took a 32-bit CPU from specification to a sign-off-clean GDS layout with no human in the loop. Treat that as a company-reported demonstration, not an independently verified benchmark — a 32-bit CPU is a well-understood design, and "no human in the loop" on a known problem is a different claim than "no human in the loop on your problem."

Founder and CEO Mark Ren is blunt about the hard part in the announcement: building production-grade agents that reliably automate real engineering work is far from easy. MediaTek's Brian Hsu frames the bet as domain specificity — the interesting product is a builder that lets engineering teams customize and improve agents for their own workflows, not a black box that claims to already know them.

Why vertical AI agents matter for your business

The unit that pays is the workflow, not the task. A tool that drafts one document faster saves minutes. A system that carries a job from intake to invoice removes the handoffs — and handoffs are where your time actually goes. When you evaluate an agent vendor, trace one complete job through it. If a human still has to move data between three steps, you bought a faster task, not automation.

Domain specificity is the moat, and you already have it. The reason a chip-design agent company can raise against general-purpose models is that the workflow knowledge, the constraints, and the proprietary data are not in the model. The same is true of your business. Your quoting rules, your exception handling, your customer history — that's the asset. A generic agent doesn't have it, and handing it to one wholesale is how you give it away.

"Customizable by your team" beats "already knows your industry." Every vertical AI pitch lands somewhere on that spectrum. The pitches that survive contact with reality assume they will be wrong about your process and give you the levers to fix it. Ask to see how you change an agent's behavior without filing a support ticket.

Discount the no-human-in-the-loop demo. Every agent company has one. It runs on a problem the team chose. The number that matters is what percentage of your jobs finish without intervention, measured on your data over a month. Ask for a pilot that produces that number.

Sources: Semiconductor Digest, Tech Startups.