Mark Ren helped prove frontier AI could do real chip-design work at Nvidia. Now his startup, Agentrys, is betting that lasting advantage belongs to companies whose engineering systems combine customer-specific workflow knowledge, continuous learning, and agent-native tools not to those that merely rent the same models.

Mark Ren has spent more than two decades working in design automation, including experience at IBM and Nvidia, where he earned an IBM Corporate Award for design closure on high-performance microprocessors, helped drive the GPU-accelerated EDA movement, and led ChipNeMo, an early large language model built for chip design, created at Nvidia before most of the industry believed such a thing was possible.

That history informs his current view: adopting frontier AI alone may not be enough to create a lasting advantage in the semiconductor industry.

"If AI is just another productivity tool everyone buys off the shelf, it raises the floor for the whole industry but gives no one an edge," Ren says.

That observation is the founding logic of Agentrys, the applied research company Ren now leads. Modern AI models can already generate RTL and testbenches, write assertions, analyze logs, and assist with debug. These are genuine engineering tasks that once required scarce human expertise. But every chipmaker can license the same capability, often from the same providers, on the same day. Productivity rises everywhere while differentiation rises nowhere. If every competitor adopts similar frontier AI, where does lasting advantage come from?

Ren's answer: from an engineering capability competitors cannot rent—the customer-specific combination of workflow knowledge, continuous learning, and accumulated operating experience, enabled by Agentrys's platform and agent-native tools.

Tasks versus workflows

Ren distinguishes between automating individual chip-design tasks and supporting broader engineering workflows. Processes such as verification, physical implementation, timing closure, analog layout, and signoff can involve multiple tools, engineers, and decisions over extended periods, with potential effects on development timelines and chip performance.

Four decades of electronic design automation automated the algorithms inside those workflows, largely through software built for human engineers. Frontier AI is now automating tasks within them. But the workflows themselves still consume enormous manual effort and judgment, and legacy interfaces often leave agents operating indirectly through scripts, terminals, and unstructured logs. Debug dominates verification time. Floorplanning gates physical design. Analog layout remains a craft.

Agentrys's thesis is that workflow automation requires more than capable models and agents. It also requires an execution layer built for autonomous systems: tools that expose design state, support controlled machine actions, preserve traceability, and return structured, verifiable feedback.

Workflows also have a property that isolated tasks lack. They repeat, on every block, every project, every tapeout, hundreds of times across a product cycle. Each run generates evidence about what worked: which stimulus found the bug, which constraint closed timing, which debug path went nowhere. In most organizations, that evidence evaporates when the project ends. Ren argues that organizations can strengthen their engineering capabilities over time by systematically capturing design decisions, actions, failures, and outcomes, then using those insights to inform future projects.

Agentrys calls the approach Agentic Design Automation, combining AI agents, engineering knowledge, and continuous learning to support broader chip-design workflows.

An AI workforce designed to compound for you

The company's stated mission is to help every semiconductor company create a continuously improving agentic engineering workforce. The underlying Agentrys platform and agent-native tools remain Agentrys technology. What belongs to—and remains under the control of—the customer is its design data, methodology, workflow history, acceptance criteria, configurations, learned adaptations, and the company-specific capability that accumulates through use.

"The agents should learn from your designs, run behind your firewall, and get better with every project," Ren says. "That is the difference between renting capability everyone else can rent and building an engineering advantage that compounds for you alone."

The demand, he says, runs ahead of the pitch. “Customers are already trying to build engineering agents using coding tools such as Claude Code and Codex. But semiconductor design requires more than individual agents. It requires Agentic Design Automation: agents, design knowledge, specialized tools, rigorous evaluation, and continuous learning working together across end-to-end workflows. Agentrys enables companies to build that capability inside their own environment—creating an AI engineering workforce shaped by their own designs, methodology, and experience.”

Agentrys Studio applies this approach within a customer’s own environment, including private-cloud and air-gapped setups. According to the company, sensitive design data and workflow information remain under the customer’s control, addressing a key concern for advanced chip-design teams.

The compounding engine is a framework the company calls Self-Evolve. Every workflow execution generates observations. Every observation is evaluated against grounded references and customer-defined acceptance criteria. According to the company, each evaluation can help refine the agents, workflows, and customer-specific knowledge. Agent-native tools support this process by providing structured information about system states, actions, errors, and results, which may help the agents adapt more effectively to a team’s designs over time.

Two further commitments complete the design. First, production flows span commercial EDA tools from multiple vendors plus decades of internal scripts, and Studio orchestrates across that mix rather than assuming a single ecosystem, operating inside the customer's existing methodology rather than asking teams to re-implement their flows. Second, agent-native tooling is foundational rather than an add-on. EDA tools were built for human engineers, with interactive interfaces and human-paced feedback. Autonomous agents need machine-actionable APIs, inspectable state, controlled actions, and structured engineering feedback that support reasoning, verification, and iteration at machine speed. Agentrys is building that layer to work alongside commercial EDA rather than replace it.

From benchmark to full flow

For a company introducing a new category, the question is whether the machinery actually runs end to end. Agentrys's answer is a published demonstration of the full thesis: an autonomous multi-agent workflow that took a small 32-bit CPU from specification to sign-off-clean GDS layout with no human in the loop, running entirely on open-source tools, with verification passing bit-exact on the first attempt and physical design closing with zero routing violations. The design is deliberately simple and the flow open source; what the demo establishes is the integrated system—agents coordinating a complete workflow, structured tool interactions, grounded evaluation, and closed loops where a physical-design decision reaches back to rewrite the RTL, which is then automatically re-verified before the gain is banked.

That end-to-end run shows the system works. A public benchmark shows how well. Rigorous, public chip-design benchmarks are scarce, which makes external validation of any AI claim hard, and verification has one of the few: Nvidia's CVDP. It gave Agentrys a shared yardstick to test its Self-Evolve method against, rather than private designs no outsider can check.

Agentrys is engaging customers today on both verification and physical design, with RTL, analog, and system design on the roadmap. Engagements follow a three-step arc the company calls Onboard, Evolve, Scale: identify one or two high-value workflows with measurable targets, integrate with the customer's tools and methodology on-prem, let the agents improve run over run, then expand across workflows and teams as confidence grows. Its customers today range from tier-1 semiconductor companies to AI-silicon startups using agentic automation to make small teams punch above their headcount, and its own team draws from Nvidia, Meta, AMD, Samsung, Google, and Siemens.

The obvious objections

The thesis invites pushback, and the sharpest version comes in three parts.

First, established design-software providers are adding AI to tools already embedded across many chip-development workflows. This raises a broader question: do companies whose platforms generate workflow data hold an inherent advantage?

Agentrys's counter is structural: the incumbents' agents are vendor-controlled and assume the vendor's own ecosystem, while real production flows span multiple vendors' tools plus internal infrastructure. An improvement loop that only sees one vendor's slice of the workflow can't compound across the whole thing, and whatever it learns accrues primarily to the vendor's product rather than to the customer's company-specific capability.

Agentrys also argues that adding an agent interface to existing software is not enough. Most EDA tools were built for human operators and batch execution, not for autonomous systems that must inspect intermediate state, take incremental actions, diagnose failures, and verify outcomes. Agentrys is therefore building agent-native tools alongside its agents: software interfaces and design utilities created specifically for machine reasoning, precise control, observability, and closed-loop execution. The advantage is not merely coordinating existing tools, but giving agents a better substrate on which to operate across the entire workflow.

Second, data thinness. Frontier models improve on internet-scale data; can one chipmaker's workflow history really move the needle?

Ren's response is that workflow learning isn't pretraining. The signal comes from repetition against grounded, checkable outcomes: hundreds of runs per product cycle, each evaluated against acceptance criteria the customer defines. Agent-native tools make that feedback more useful by exposing structured design state, actions, errors, and results rather than leaving the system to infer progress from terminal text and log files alone.

The idea is not that one company’s data creates a stronger general-purpose model, but that agents may become more effective within that company’s designs, methods, infrastructure, and tools. Over time, customer-specific configurations, knowledge, and adaptations may also become more useful.

Third, erosion from above. Task-level AI advantage has evaporated with every model release; why would workflow-level advantage survive the next one?

Here Ren's answer is close to the thesis itself: better base models lift every task inside the workflow, for everyone, simultaneously. What may be harder to replicate is the accumulated knowledge of a specific organization’s designs, decisions, setbacks, acceptance criteria, and engineering methods. That information remains within the customer’s environment and under its control.

Nor does a stronger model automatically create the execution environment required for dependable design automation. It still needs access to tools that expose the right state, support controlled actions, preserve traceability, and provide verifiable feedback. Agentrys argues that its approach combines customer-specific workflow knowledge, a learning system spanning the design process, and tools that make engineering work easier for AI to interpret and execute. Stronger models may improve performance, but they may not replicate the capabilities developed within a customer’s own environment.

Whether those answers hold is the question the current customer engagements will settle. What the crowded field confirms, at minimum, is that AI in chip design is no longer contrarian. EDA vendors, foundation-model startups, and design houses assembling internal agents with tools like Claude Code and Codex are all converging on the same territory.

Ren’s contention is that most are still competing at the task layer, where advantages erode as models improve for everyone. The more durable prize sits one layer up: a customer-owned system that learns across complete workflows and one layer down, in the agent-native execution tools that allow AI to interact with chip design environments with far greater precision, visibility, and reliability.

What compounds may gain momentum

Ren suggests that while AI and task automation may become widespread across the semiconductor industry, companies that combine internal workflow knowledge, continuous learning, and agent-native tools may develop stronger engineering capabilities over time.

It's a thesis with home-field resonance in an industry whose entire history is a story of compounding, from transistor density to process learning to design reuse. Ren is betting that the engineering system itself is the last great thing left to compound. Agentrys supplies the platform and agent-native tools; each customer builds the proprietary knowledge, adaptations, and operating capability that can make its own system pull away.

VentureBeat newsroom and editorial staff were not involved in the creation of this content.