Skip to content
Theme
OnDevice Tools

Tech BusinessChips & Semiconductors

AI agents emerge as reusable design tools for chipmakers

Chip design firms are exploring AI agents that can customise and optimise workflows faster than traditional IP blocks, but trust, verification and pricing models remain unsettled.

Illustration of AI agents working on a semiconductor chip design layout with verification checks

Key takeaways

  • AI agents are being positioned as reusable design tools that can customise chip workflows faster than traditional IP blocks.
  • Verification and debugging by experienced engineers are still required before agent output can be trusted.
  • Orchestrating multiple agents that run in parallel is a major technical and business challenge.
  • EDA vendors are discussing consumption-based and token-based pricing, but adoption remains limited.
In this article
  1. Agents as a new kind of design IP
  2. Faster customisation, but verification remains
  3. Orchestration becomes the hard problem
  4. Pricing models are still taking shape
  5. What it means for chip designers

Agents as a new kind of design IP

Illustration: code, AI, price, crypto and the labels AI Agents, Eda, Semiconductor Engineering

AI agents are being pitched as a fresh category of reusable design capability for semiconductor companies, according to Semiconductor Engineering. Unlike fixed intellectual property blocks, agents can explore design options, integrate changes and refine results with little human input. Supporters say this could make customised design workflows quicker and cheaper to build.

The comparison with soft IP is deliberate. For more than a decade, soft IP developers have tailored code for large customers to improve performance or cut power. That work is usually narrow and expensive, and gains for one chip may not carry over to derivative designs, particularly multi-die assemblies at advanced nodes.

Faster customisation, but verification remains

William Wang, chief executive of ChipAgents, said his company runs in-house agents for tasks such as specification analysis, register-transfer-level generation, formal and functional verification, clock-domain-crossing checks and place-and-route. He said some customer work resembles the IP business, but the cycle is much shorter because agents are easy to customise, with some solutions delivered the same day.

Those outputs still need to be verified and debugged by experienced engineers. The attraction is speed: agents can sort through many options quickly as AI models evolve. Mo Faisal, chief executive of Movellus, compared the current period to earlier architectural shifts, saying the industry goes through similar churn every 10 to 15 years.

Orchestration becomes the hard problem

Moving from generative AI to agentic AI adds a coordination layer. Orchestration defines how tasks are assigned, how agents reach tools and data, and how work is kept in order, especially when several agents run at once and their actions overlap.

Robbie Williamson, who leads customer engineering for cloud AI at Arm, said the difficulty is not a wholly new workload but coordinating many parallel agents at the CPU level with enough bandwidth and memory to run at scale. Prith Banerjee, senior vice president of innovation at Synopsys, described a similar hierarchy in chip design, where agents for EDA and computer-aided engineering tools must balance competing goals such as power and area, much as human teams do.

Pricing models are still taking shape

Business models around agents have yet to settle. Rob Knoth, senior group director of strategy and new ventures at Cadence, said the EDA industry moved from perpetual licences to time-based licences years ago. Agents are pushing tool usage higher because they run many experiments, which makes consumption-based pricing attractive to customers who already understand it.

Wally Rhines, chief executive of Silvaco, said he is open to selling an agent that automates calibration and draws on simulated data, with billing based on tokens consumed rather than access to proprietary data. He noted that customers are discussing token-based pricing even though no EDA vendor appears to have adopted it widely. Security and permissions for agents are also not always well defined, which may slow adoption.

What it means for chip designers

The shift matters because it could change how design teams buy tools and organise work. Agents promise faster exploration and customisation, but they also require orchestration, verification and clear rules for data access. Companies will need to decide whether to build agents in-house or buy them, and how to measure the value of token-based usage against traditional licences.

For now, the technology is emerging rather than established. The sources describe interest and early deployments, not a settled market. Design teams should expect a period of experimentation, with engineering oversight remaining essential before agent-generated results are trusted in production chips.

Sources

  1. Are AI Agents The New Semiconductor IP?Semiconductor Engineering

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.