A single mistake at 3nm or 2nm now costs roughly $500 million in lost wafer value and recovery work, according to Ankur Gupta, executive vice president and head of EDA IC software at Siemens EDA. That figure is changing how the largest electronic design automation vendors structure their software. They are moving away from single large language models and toward fleets of smaller, task-specific agents that work in parallel under human direction.

Why move from large models to small ones?

Large language models can spin up dozens of agents for open-ended reasoning, but that generality is wasteful for chip design. PPA optimization, floor-planning, verification, and yield analytics each benefit from a narrower model trained on a narrower data set.

"In our early work on agents, which we created for each of our tools, we have questions and answers," said Prith Banerjee, senior vice president of innovation at Synopsys. "We had an agent for this, and an agent for that. But in this world of multi-agent workflows, our customers are trying to design a chip. Maybe they're having a problem with the PPA, and some engineers say they know exactly what changes are needed to lower the power."

The principle is division of labor. Each agent holds a limited scope. A coordinator above it stitches results together.

"The model is drawing a boundary," said Mo Faisal, chief executive of Movellus. "It's like, 'Hey, you're getting in the domain that I work in. Stay in your lane.' We will specialize. A couple of large companies will take the LLMs, but everybody else has to specialize."

Takeaway: chip design is shifting to small, scoped agents under a coordinating layer, with humans directing.

What does the new toolchain look like?

The largest vendors are now selling what Cadence calls super-agents. These software constructs orchestrate sub-agents for specific design tasks.

"We've gone to the aggregation level with super agents, and we think their complexity and the amount of value they can deliver is as big or bigger than EDA tools," said Matt Graham, senior group director of verification software product management at Cadence. "These are a new category of tools, and we're starting to talk less about agents than skills that are dynamically assembled into agents."

ChipAgents shipped its own multi-agent coordination package for debugging and verification in the fourth quarter of last year, said chief executive William Wang. Shriniwas Kulkarni, an engineering leader at SiMa, reported similar momentum in retrieval-augmented generation workloads.

Scheduling and message-passing between agents often demand more compute than the matrix multiplications themselves. Several executives noted that the orchestration problem is now harder than the inference problem.

Takeaway: orchestration of specialized agents is the central engineering problem, not the underlying models.

How do vendors fit agents into existing tools?

Every major EDA platform has a graphical interface built around an engineer clicking through steps. Agents require something different.

"Typically, customers will work within a platform, and that could be Cadence Virtuoso, Synopsys Custom Compiler, or Keysight Advanced Design system, and within that environment we guide them through an application-specific workflow," said Niels Faché, senior vice president at Keysight EDA. "All of that coding is typically done in Python, and that would then allow an agent to stitch together a code-based workflow. Our job is to make sure that our customers get this MCP functionality."

Vendors are wrapping their tool internals in Model Context Protocol servers so an external agent can call them like a function. Anthropic introduced MCP as a standard. Keysight and others are moving to expose every step from synthesis to sign-off.

Takeaway: the EDA tool stack is being rebuilt around API-callable functions rather than point-and-click menus.

What is the human role?

The talent shortfall makes the question acute. "We went through the wonderful dot-com period where nobody went into semiconductors because they all wanted to be software programmers," said Michal Siwinski, chief product and marketing officer at Arteris. "There was a lull in the talent pool, which has been recovering slowly, but there's still a gap in terms of the number of experienced people who actually impact designs. Agentic AI helps to fill some of those voids. In this case, it's not actually an accelerant. It's backfilling."

Every executive in this story agreed on one thing. Humans stay in the loop, and they stay as the interpreter when an agent overreaches.

"If you have a black-box view of the entire problem, do you build some black box that is hierarchical to get the best result, or do you want to pause as you go through this process and have a human understand so they can bake into it," said Saurabh Gayen, chief solutions architect at Baya Systems.

"AI isn't so much a replacement for engineers as a way to enhance what they're able to do and their understanding," he added.

Gayen also warned against early certainty. "How you engineer those layers of agents is an art. And if somebody says they've found the perfect solution, that's really just one among many."

Takeaway: agentic AI is filling missing expertise, not replacing the engineers who set goals and approve results.

How might EDA pricing change?

Silvaco chief executive Wally Rhines expects token-based pricing to enter the conversation. "It's not clear if anybody in EDA is actually pricing based upon tokens rather than traditional pricing models," Rhines said. "I don't think anyone is, but their customers are all talking about it."

Such a model would let a vendor meter work by tokens consumed while keeping a customer's fab data local. The agent would run on synthetic data calibrated against physical measurements.

"The fabs aren't generating these AI models yet," Rhines said. "Everyone wants to use Claude for everything."

Takeaway: token-based pricing is on the table, but no EDA vendor has shipped a public price list built around tokens.

What comes next?

The executives sketched a path in which orchestration skills accumulate faster than the language models underneath them.

"With multi-agent coordination, in the future we're going to have virtual chip companies that are also going to be virtual AI engineers working together and collaborating to form this agent organization," said ChipAgents's Wang.

Movellus's Faisal drew a parallel to past transitions. "You can draw parallels to the early 2000s, when the gigahertz wars all of a sudden became multicore," he said. "The bottleneck will shift to data movement, and then it will move to something else once we solve that."

Within two to three years, expect EDA vendors to ship MCP-wrapped toolchains with priced-by-token super-agents, while the engineers who remain spend more time setting constraints and checking outputs than running scripts by hand.