Cadence told an audience at Computex on June 1 that its ChipStack agent has reached what it calls Level 5 autonomy in a single domain — the boldest claim yet in a year in which the three dominant EDA vendors, Cadence, Synopsys, and Siemens EDA, have all shipped agentic AI tools for chip design. The caveat matters: Rob Knoth, Senior Group Director of Strategy and New Ventures at Cadence, described the demo as "what we call bounded Level 5 in one domain," with most of the super-agent technology at "advanced Level 4."

Synopsys, which showed a spec-to-RTL workflow on March 11 rated "L4," now claims L5 capability for its long-horizon agents. Siemens launched its Fuse EDA AI agent on March 16 and announced "self-verifying" loops on July 26. All three agents pair a reasoning model — mostly Nvidia's Nemotron running in the OpenShell sandbox — with deterministic EDA engines such as Xcelium, Jasper, Questa, Calibre, and Fusion Compiler, which validate every proposal the model makes inside an iteration loop.

The headline numbers concentrate in front-end design. Synopsys claimed in July that RTL validation could run up to 50x faster with 20% better coverage versus its own non-agentic flow. Cadence said in June that ChipStack, in early access since February, can cut "a typical five-week verification loop to less than a day" — over 40x — with each engineer running "hundreds of dynamic simulations with Cadence Xcelium." Named early ChipStack users include Altera, Nvidia, and Qualcomm; Altera cited about 10x improvement "in some areas" of verification effort in Cadence's February release.

Every one of these figures is vendor- or customer-reported, often with "up to" or "early evaluation" caveats, and each measures something different against a different baseline. The only evaluator Synopsys named in July, AMD, has offered an endorsement but no independent analysis.

The autonomy ladder is marketing scaffolding

Cadence rates its agents from optimization AI through conversational LLMs, complex reasoning, and agentic workflows up to full autonomy — spec to verification "with minimal human intervention." Synopsys' ladder runs from single-agent, single-step actions toward fully autonomous decision-making, though the company stresses that "human engineers will and must always be in the loop." Siemens skips numbered levels entirely. Behind the different labels sits the same architecture: an agent that proposes, a deterministic engine that checks, and a loop that runs until verification closes.

Siemens' Questa One Agentic Toolkit, detailed in February, ships five separate agents — RTL Code, Lint, CDC, Verification Planning, and Debug — and works with mainstream coding applications including Claude Code and Cursor alongside Siemens' own Fuse. Akshay Aggarwal, senior director of engineering at MediaTek, reported "immediate and significant" gains.

The back end resists the agents

Implementation is a harder problem. Physical design deals with geometry and physics, not code, and runtimes are compute-bound. Tool calls that take "hours to days" to return a value give "the AI industry a much different problem to solve," Knoth said. Claimed gains thin out accordingly. Cadence's InnoStack runs parallel experiments from synthesis to ECO execution; ViraStack, aimed at custom and analog design, has some customers "reporting 3–10x productivity improvements" according to an April Futurum report. Siemens said its Solido tools cut characterization turnaround by more than 10x with 5x-to-10x lower token cost, and STMicroelectronics' Gianbattista Lo Giudice said the Solido Layout Analyzer helps "cut down the time we spend debugging complex design blocks by weeks."

China's Empyrean, the largest domestic EDA vendor and state-controlled since December 2024, is building its own platform. Chairman Liu Weiping said on Sept. 9 that its agent cut one layout task from four weeks to one, per the South China Morning Post, and TrendForce reported 3D verification dropping from two-to-three months to under a month. Liu said Empyrean is moving from "humans operating tools" to "humans commanding agents," with a shift from licenses toward token-consumption pricing.

Trust, guardrails, and no general availability yet

Self-checking remains the core trust question. "The crucial approval checkpoints are still going to be human-driven," said Anand Thiruvengadam, executive director of product management at Synopsys, though he noted customers over time "might relax and take away those checkpoints." Nvidia's newly announced Open Agent Safety Platform — OpenShell plus a Sentry watchdog on BlueField-4 DPUs that can quarantine an agent in milliseconds — counts all three big vendors among more than 100 participating organizations.

Nothing is generally available yet. Cadence's Level 5 early access is due in the second half of 2026, alongside its AuraStack packaging agent; Synopsys promises general availability of its seven-agent AgentEngineer portfolio — with 50-plus engagements — by end of year; Siemens' self-verifying releases are forthcoming. The differentiator to watch through year-end is whether any vendor names a customer running a top-level agent in production, on real workflows, with measured results.