DAC 2026: autonomous agents arrive in chip engineering
Three pillars reshaping semiconductor design
DAC 2026 revolved around three themes that will dominate the coming years: agentic AI as a new automation layer, computational physics as a fundamental bottleneck, and open standards as the glue that makes all of this scalable. These are not minor changes. The industry has moved from talking about agents in demos to real deployments in customers' hands.
The first step is understanding what "agentic" means in this context. It is not a chatbot that answers questions. These are long-running autonomous workflows that orchestrate verification, thermal analysis, and implementation tools, parse log files, decide what to try next, and keep context across weeks of continuous execution. Ravi Subramaniam, head of product at Synopsys, sums it up this way: the industry has moved from brief chat-driven interactions to workloads that last days, which lets agents keep real autonomy without losing the thread of what they are doing.
Tim Costa, VP at Nvidia, frames it from the engineer's angle: today the limit in product development is not how many experts you can hire, but how much compute you can dedicate to a chain of agent-orchestrated tools. In theory it frees engineers from repetitive tasks and returns them to what brought them to engineering: thinking about systems, not how to use a specific tool. A motorsport engineer wants to think about speed and aerodynamics, not the buttons of a software package.
Physics as the frontier
But agents are only as good as the tools and models they orchestrate. That is why computational physics is the second leg of the triangle. Startups like Vinci train physics models that aim to replace weeks of simulation with automated analysis behind the corporate firewall. Founder and CEO Hardik Kabaria puts it bluntly: "We are the ChatGPT of physics, but deployed in production, not as a hobby." He claims a three-orders-of-magnitude speedup on real package analysis versus current flows.
The problem they solve is well known: from a $4 trillion hardware economy, there are only about a million engineers who truly know how to do physics. Kabaria adds that physics does not disappear — it is democratized. On the analog design side, Maieutic Semiconductors uses "Socratic" AI that questions engineers and mines internal data to extract the tribal knowledge of senior designers. As co-founder Ashish Lachhwani says, "in digital there was automation and shared data years ago. In analog, a five-transistor circuit can take years to perfect."
Silvaco, under new CEO Wally Rhines, sees the revolution from a different angle: autonomous agents that act as democratizers between fabs and designers. Instead of forcing a fab engineer to become a TCAD expert, an agent can query digital twins in natural language, run large-scale design of experiments to diagnose yield anomalies, and optimize process steps in real time. As Rhines puts it, "agents are the great equalizers. They look for the best solution, gather the best data. If I had to be a TCAD expert, I would have to go to the TCAD team. Now I have an agent with an English interface that can look it up for me."
The real bottleneck: standards and interoperability
Intel, through Lalitha Immaneni, offers a clear perspective: design teams still often jump "blindly" into complex stacks without multi-domain physics guidance or standardized data. Malformed copper density data forces fab mechanical teams to spend three or four weeks cleaning the database before another three or four weeks of analysis. "That is not acceptable at this complexity," she says.
Intel's answer is system-technology co-optimization (STCO) based on packaging design kits (PADKs), the packaging equivalent of a PDK. PADKs bundle design rules, reference flows, training materials, and Intel-specific verification scripts for technologies such as EMIB and TSV. But the key point is that verification cannot be late: you have to validate and model from the start.
True scale requires open standards. Lu Dai, president of Accellera, explains his organization's role: publishing machine-readable, interoperable specifications. One of them is the draft Functional Safety Language (FSL) 1.0 standard, which places functional safety intent between tools and teams automatically and traceably. FSL is metadata, like the Unified Power Format (UPF), so tools can reason about safety requirements and coverage directly.
Without interoperability, no realistic agentic flow can cross foundries, IP providers, and tool vendors. Paul Penzes, VP of engineering at Qualcomm, puts it plainly: "We cannot afford a closed system because we will not succeed. In mobile every millimeter counts and we need an open system where we can experiment to get the best chip."
The business-model shift
Prith Banerjee, SVP of innovation at Synopsys, asks a question that defines the challenges ahead: in a world of tools from different vendors, will there be an open agentic framework? The industry's answer is yes, but that forces business models to change. Selling perpetual or annual licenses does not work with token-consumption-based flows. If an agentic flow burns tokens based on usage at your site, a flat price makes no sense. The model has to be real consumption, but the industry has not yet articulated this clearly.
Startups such as Agentrys and Normal Computing explore new layers. Agentrys, led by Mark Ren (former director of design automation research at Nvidia), proposes that the design artifact of the future will not be RTL but the agents themselves, redefining the design automation stack. Normal Computing pairs EDA tools with "unconventional" hardware such as analog matrix engines and stochastic chips, aiming at two-orders-of-magnitude efficiency gains.
Why this matters for conversational hardware
At bitbitbla the agentic brain of our platform does something that anyone who understood the article will recognize: it processes user intent, reasons, makes decisions, and identifies context from the environment. It crosses audio transcription with the current state of devices to resolve ambiguous commands — "turn that off" in the right zone — and uses RAG to enrich intent with technical documentation and run complex commands. From the Dashboard you define the agent's personality, the tools it uses via MCP, the JSON payloads and webhooks that connect to the customer's infrastructure, all by remote configuration without reprogramming the hardware.
What is happening at DAC 2026 in chip engineering is the same pattern: agents that understand multi-domain context, that reason about constraints and state, that make autonomous decisions while keeping user intent as the compass. The scale difference is huge, but the problem is analogous. That is why the EE Times article builds a double bridge: not only is agentic AI the new engine in EDA, but any embedded system that must make complex decisions on-device is navigating exactly the same waters.
Source: EE Times.
