Synopsys puts 30 agents inside chip design — and bills by usage
Synopsys used its Shanghai developer conference to answer the question every EDA customer is now asking: when agents do the engineering work, what exactly are you buying? Also today: Singapore's digital minister says AI will need aviation-style safeguards.
Synopsys chief executive Sassine Ghazi told the company's developer conference in Shanghai that AI agents, co-design and electronic digital twins are the three investments that will define engineering over the next decade — and that Synopsys already runs more than 30 task-level agents, each an expert user of one tool or one step in the flow. The agents are organised in layers: a single agent handles a defined task, several of them working together form a domain-level system — in verification, one agent locates the bug, one judges whether the tests cover enough, one generates or adjusts tests, one runs the tools, and an orchestrator above them sequences the work — and long-running agents sit above that, reasoning across engineering domains for hours at a time. Synopsys' stated limit is context, not capability: an agent is only as good as the design rules, historical project data, proprietary IP and quality standards it is given.
In a briefing alongside the conference, Ghazi said the definition of a chip design engineer — and of a chip design company — changes within one to two years. In his version, a system company writes the specification, generates RTL with a model, and hands that RTL to a design house for verification, physical implementation and manufacturing preparation. Agents may keep pushing further down that chain. What does not change, he said, is the sign-off checklist: whether humans designed the chip, agents assisted, or agents did most of it, the same deterministic checks and manufacturing, performance, reliability and safety standards apply before anything is built.
The commercial detail is the part worth marking. Ghazi described two agent businesses: a multi-year subscription to Synopsys' own agents, and software consumption — agents calling Synopsys' EDA tools meter out usage the way a seat license never did. He said customers building their own agents are not competition but evidence of a new business model, and that the platform will stay open to third-party agents. Synopsys reports doubled token efficiency and better silicon quality in joint agent work with Nvidia, AMD and Microsoft shown at this year's DAC, and its first Multiphysics Fusion products — from the Ansys acquisition that closed in July 2025 — are already deployed with customers. It holds an investor day on September 30 to say more about how agents get monetised.
Our read: the billing change is the real announcement. A verification run that fires thousands of tool invocations converts EDA from an annual seat-license annuity into a metered one, and the doubled-token-efficiency pitch is what makes that sellable to customers who will otherwise watch the meter. The moat Ghazi is describing is not the agents — anyone can build those — but the sign-off checklist they have to satisfy. That is where the margin lives, and it is the one part of the flow Synopsys says it will not hand over. Hardware is becoming the next serious eval surface for agents — we covered the first end-to-end version of this in AI designs, verifies and deploys a chip in two weeks.
Singapore's minister for digital development says AI will eventually need the kind of regulatory scaffolding civil aviation took decades to build, and that the country is deliberately building its safeguards alongside adoption rather than after it. In a LinkedIn post, Josephine Teo pointed to the "hundreds, if not thousands" of standards covering aircraft manufacture, maintenance, airport and airspace management, and said trust is what made people board planes: "Trust is an essential requirement for a technology to be regularly relied upon." She named AI Verify, the Singapore AI Safety Institute and the city-state's model AI governance frameworks as the practical work already underway, and conceded that safeguards for advanced models still need research and testing — "this takes time." Singapore also hosted May's International Scientific Exchange on AI Safety, which drew more than 100 contributors from 13 countries and produced the Singapore Consensus on global research priorities.
Why it matters: aviation is doing double duty in that analogy. It is the argument for regulation, and it is also the argument for patience — the industry grew into its safety regime, and no one wrote it up front. That suits a country that competes for AI investment on trust rather than compute, while the frontier labs argue for self-regulation and Anthropic pushes a coordinated slowdown. The tell is the mechanism Teo keeps returning to: accreditation of third-party testers countries can point at, not rules that stop a training run. Whether "trust" is a durable differentiator depends on disputes those testers have no power to settle.
What to watch: whether Synopsys puts an actual price on agent consumption at its September 30 investor day, and whether the first US lab adopts aviation-style incident reporting before Singapore's testers get anything to test.
If agents do the design work and the checklist is the last human gate, what is an EDA company actually selling — the tools, the agents, or the sign-off? Tell us in the comments.
Sources: 智东西 Zhidx · 电子创新网 EETrend · Sohu — 2026 Synopsys Developer Conference · Synopsys — Agentic AI chip design with AMD and Microsoft · Synopsys — Investor Day 2026 · CNA · The Straits Times · AsiaOne · MDDI — Minister Teo at the International Scientific Exchange for AI Safety
