Vendors are pitching AI agents to draft behavioral models (engineering shortcuts that approximate a block's input output behavior without re running the full physics) for analog and mixed signal (chips that mix real world analog signals with
For analog and mixed-signal chips, running full transistor-level simulation across every corner can swallow weeks. Sometimes months. A behavioral model is the engineering approximation a team uses in place of re-running the full physics every time a downstream simulation needs to touch that block: the input-output behavior without the internal detail. Agentic AI vendors are pitching workflows to draft that model. The harder truth, even the source material concedes, is that "good enough" is still a human judgment call.
The blocks in question are the awkward ones. PLLs, the on-chip clock generators, have to behave across jitter, supply, and temperature. ADCs and DACs, the analog-to-digital and digital-to-analog converters at the chip's interface with the real world, vary across input range, noise, and process. LDOs, the low-dropout regulators that hold the power rail steady, swing across transient load, dropout, and stability. Running full transistor-level simulation on any of them across every process, voltage, and temperature corner is exactly the kind of compute bill that justifies a behavioral shortcut in the first place.
The most detailed walkthrough, a SemiEngineering-hosted explainer from ChipAgents, lays out a six-step agentic workflow. The vendor's own page describes a multi-agent platform built around gap detection, stimulus generation, active learning, and model refinement. The loop, in either telling, runs the same way: define scope and requirements, generate reference data, produce the model, validate against golden simulation, refine, and deploy. The output lands in Verilog-A, SystemVerilog, C++ DPI, or a vendor-proprietary format.
The agent can take the mechanical layer off the engineer's plate: spinning up reference data, detecting which corners of the input space the model misses, generating stimulus to cover them, and running the draft through the validation harness, which compares waveform shape, timing, frequency response, power, transient behavior, startup and reset, and corner-case response against acceptance criteria. The engineer still owns the judgment layer. Which corners of process, voltage, and temperature actually matter for the next integrator. How much deviation from the golden waveform is tolerable before the model misleads. How you prove "good enough" in a tape-out review. Three of the six steps are judgment calls an engineer has to defend: scope, tolerance, and sign-off. The other three are candidates for autonomy. The boundary between them is the actual product.
This is the niche where the wider agentic-EDA race is landing first. Cadence launched ChipStack as its own agentic offering for chip design and verification, and The Register covered the rollout in February 2026. Synopsys has been pushing a longer arc: by February 2023, the company had counted 100 commercial tape-outs using its AI-designed chip technology, and the Q2 FY2026 earnings call was explicit that AI-driven design demand was a tailwind. The shift from AI-assisted to agentic tooling is happening in this corner of the workflow first because the simulation bottleneck is most acute there, and because the loop is procedural enough to script without pretending to replace the engineer.
The current source cluster, on its face, is a how-to blog from one vendor, hosted on an industry publication, plus three vendor press releases and two industry-press pieces. No independent AMS designer's testimony sits in the bundle, no second startup's parallel effort is named, and no measured productivity claim, no runtime number, no tape-out outcome, no customer count survives scrutiny. The follow-up worth chasing is straightforward: which design teams have run a behavioral-modeling agent end-to-end on a real PLL or ADC, where the model broke, and what it cost in extra validation cycles to fix. Until that reporting lands, the defensible claim is narrower than the vendor pitch. Agentic AI is changing how the modeling loop is staffed, not who owns the judgment that gates the model.