Will AI Agents Become the Next Generation of Semiconductor IP?
AI agents are emerging as a new class of reusable capabilities in chip design. ChipAgents CEO William Wang says traditional IP development can take 10 months, whereas custom agents can deliver a solution in some scenarios within a single day. Agents can automatically complete specification analysis, RTL generation, formal verification, and place-and-route, but engineering teams are still needed to verify the results, and the business model as well as data sharing and trust verification issues have yet to be settled.
Body · AI translation
Will AI Agents Become the Next Generation of Semiconductor IP? Source: Semiconductor Horizons Published: 2026-11-16 17:32 AI chip manufacturing generates poster. Traditional IP development may take 10 months, while the customization cycle for AI agents is much shorter—in some cases, a solution can be delivered to customers the same day.
AI agents are becoming a new class of reusable design capability, but they differ from traditional IP. Agents do not provide design modules with fixed definitions; instead, they can be used to explore multiple options, complete design integration, and iteratively optimize designs based on these. At advanced process nodes, soft IP vendors have been customizing code for major customers for more than a decade to improve performance or reduce power consumption. However, such customization is usually limited in scope and expensive, and optimizations made for one chip often cannot be directly reused on derivative chips—especially for multi-chiplet packaged products at advanced nodes. In contrast, agents can customize different functional modules with little or no human intervention; they can also automatically optimize for different workloads and application scenarios, and some agents can even respond to changes caused by hardware aging. Although the final results and recommendations still need to be verified by experienced engineering teams, agents complete this work far faster than human engineers. "We have dedicated AI agents internally, each with its own role, used for specification analysis, RTL code generation, formal verification and functional verification, as well as CDC clock domain crossing checks, place-and-route, and other work," said William Wang, CEO of ChipAgents. "We deliver optimized agents to customers, and some customers build their own AI agents. It is similar to the IP business: sometimes we customize a dedicated agent for a single customer for a specific flow, which is very much like the IP business. The difference is that our delivery cycle is extremely short. Traditional IP development may take 10 months, while the customization cycle for AI agents is much shorter—in some cases we can deliver a solution to customers the same day." "This reminds me of the GHz clock speed race in the early 2000s, when the industry suddenly shifted entirely to multi-core architectures," said Mo Faisal, CEO of Movellus. "Before that, everyone thought all chips would be GPUs, but later it turned out not to be the case—CPUs and GPUs are used together, and the ratio is not fixed at 1:1. In fact, we are repeating a major architectural iteration that occurs every 10 to 15 years." The business model for agents has not yet taken shape, and it needs to be flexible enough to keep up with the pace of technological iteration. "Looking back at the development of the EDA industry, we stopped selling perpetual licenses long ago and shifted to time-based subscription licensing many years ago," said Rob Knoth, Senior Director of Strategy and New Business at Cadence. "Now that agents are emerging, the same trend is appearing. Agents call tools extensively, conduct experiments frequently, and try and fail quickly. Usage-based billing is more suitable, and customers find this payment method easier to understand." But the transition is not simple. Although agents have broad prospects, their adoption in chip design still takes time. In addition, the definition of security permissions is not yet clear. "In some scenarios, selling agents directly is a better choice than selling software," said Wally Rhines, CEO of Silvaco. "Many people are watching to see whether EDA vendors will stop using traditional pricing and switch to token-based billing. So far, no vendor has done so, but customers are all discussing this direction. I would be very willing to sell an agent externally to automatically complete manufacturing equipment calibration; the agent can compare simulated synthetic data with measured physical data. We cannot see the customer's raw physical data, but we can count the number of tokens consumed by the agent. Under this model, EDA vendors can deliver products and earn revenue, and the customer's core confidential data will not be leaked." Divide and Conquer, but Hidden Risks One core change in moving from generative AI to agentic AI is the introduction of orchestration capability, which will have far-reaching business implications. Orchestration is essentially a top-level scheme that defines how tasks are allocated, how agents call tools and data, and how various work items coordinate. If multiple agents execute in parallel and their tasks overlap, the difficulty of orchestration increases substantially. "The difference between traditional AI and agentic AI lies in how to coordinate multiple agents working in parallel and undertaking different tasks," said Robbie Williamson, Vice President of Cloud AI Customer Engineering at Arm. "It is not that unprecedented new workloads have appeared. The core difficulty is how to coordinate scheduling at the CPU level, ensure sufficient bandwidth and memory resources, and support large-scale operation. It is no longer a human initiating requests and driving AI computation; instead, programs complete it automatically, much faster. In our conversations with hyperscale cloud providers, we found that what they care about most now is what capabilities the CPU needs to provide for this. This is less prominent during model training, where training just submits tasks in batches and waits for results to return."
This creates a new hierarchy in the design flow. "After multi-agent workflows are deployed, companies will encounter scenarios like this when carrying out chip design:" said Prith Banerjee, Senior Vice President of Innovation at Synopsys. "When PPA metrics fall short, some engineers will propose using clock gating to reduce power. But after enabling clock gating, other problems arise, such as an increase in chip area. This is exactly the same coordination challenge encountered when multiple engineers co-design a system. In a multi-agent framework, the work originally coordinated by a human supervisor among engineers is now completed by a hierarchical agent system. We have agents adapted to EDA tools and agents adapted to CAE computer-aided engineering tools; the simulation module includes three major solvers: computational fluid dynamics CFD, electromagnetic simulation HFSS, and structural mechanics simulation. In the EDA domain, there are agents corresponding to the implementation stage and the signoff stage respectively. Under each type of tool, directions are further subdivided, each handled by a corresponding agent. The work originally coordinated and completed by an engineering team and supervisors is now handed over to this hierarchical agent system."
This will subtly change the working model of chip design. Soft IP can usually migrate across process nodes, but generally requires re-optimization and verification. AI agents can coordinate EDA tools, compare multiple options, automatically complete this type of work, and can also work with traditional IP to optimize specific functional modules within a complete large design.
"One of the core impacts of agent workflows is decoupling," said Niels Faché, Senior Vice President of the Design Engineering Software business at Keysight. "We have always adhered to a platform approach. Now the industry needs to think about how to adapt to the decoupling trend. The platform itself has strong user stickiness; many customers invest heavily in customizing the platform, making it hard to migrate, and the same is true for Cadence and Synopsys. But we believe customers will ultimately pursue flexibility. We choose an open route and embrace the decoupling trend. We need to leverage our strengths in simulation capabilities and industry understanding. Customers who want to work with us to build workflows need support for orchestration and scheduling capabilities, so we need to go deep in this area; while customers with top technical capabilities will complete orchestration on their own."
This requires EDA vendors to make fundamental changes. "The model changes after introducing AI; engineers will work alongside digital agents," Faché said. "So we need to restructure our products. You have probably heard of MCP (Model Context Protocol) servers. We are re-examining our products: how to modularize them and expose various capabilities through standardized APIs? Most of the relevant code is written in Python, so agents can string together complete workflows based on the code. Our goal is to let customers use MCP capabilities while also providing examples so agents can learn from them. When a new design requirement is received, the agent can organize tasks based on the experience it has learned. This is the direction we are heading, and leading customers are already building this type of agent workflow, which also requires us to provide MCP servers and complete documentation to train agents."
What other capabilities do agents have?
How much autonomy agents can have is still undecided. AI has black-box characteristics and, unlike traditional IP, requires continuous observation of its operating state. When multiple agents are responsible for different but interrelated tasks, complexity rises further.
"Orchestration and scheduling will be a major challenge, and it will keep getting harder," said Stelios Diamantidis, Chief Product Officer of Cognichip. "Agent capabilities continue to improve, and feasibility has been validated in many areas. But grouping a large number of agents to collaborate and jointly solve the same problem will spawn many new problems. Most current models do not understand the underlying physics behind chip design, and a large amount of logic needs to be built into the agents. What happens once different agents give contradictory conclusions? Will the EDA domain see back-and-forth tugging that cannot converge? These problems are not unfamiliar to us."
This problem is especially prominent in multi-chiplet packaging scenarios. "Stacking solutions are highly heterogeneous," Diamantidis said. "The data sources for training models are often of unknown origin. Different teams feed large amounts of logic into agents, and when multiple agents and models are combined, the overall system becomes very chaotic. Once stacked, uncertainty is amplified exponentially."
These challenges also explain why IP and EDA vendors are deeply integrating agents into existing tools.
"We apply agents in data center design and operations tools; agentic AI combined with physical simulation can integrate massive heterogeneous data, initiate queries, run simulation experiments, evaluate results, and then iterate in a loop," said Knoth of Cadence. "Our industry has a unique advantage: unlike other industries, there is no need for manual judgment of whether results are right or wrong. The tools accumulated in EDA over decades can themselves judge whether a solution is good or bad. For example, this layout solution is better and has lower power; or this layout performs poorly, with both power and heat worsening."
Building an agent hierarchy is important, but implementation is not simple. "We can use one main agent to schedule multiple sub-agents to divide the work and collaborate, forming a multi-agent solution," Wang said. "We released a multi-agent solution last year. SoCs are enormous in scale, and it is very difficult for a single agent to complete performance tuning; during the debugging phase, multiple agents need to read different contexts in parallel to carry out work. The same is true for formal verification scenarios: multiple agents debugging synchronously and optimizing the design is very critical, and multi-agent collaboration must be done well. In the future, virtual chip companies will emerge, with virtual AI engineers and agent clusters collaborating to complete designs, and part of the business form of chip companies will move toward virtualization."
This in turn will also accelerate the deployment of agents.
"In the future, we may model agent workloads before tape-out in order to optimize the supporting IP, while in turn influencing the CPU architecture and the CSS products we offer to our custom customers," said Williamson of Arm.
Trust and Verification
The bigger controversy lies in how much autonomy chip designers are willing to give agents.
"At the 3nm and 2nm nodes, the loss from a single design failure can reach $500 million," said Ankur Gupta, Executive Vice President of IC Products at Siemens EDA. "So this type of intelligent engine must have determinism and high fidelity. Of course we are using reinforcement learning and unsupervised learning, and have been doing so for more than a decade. But the criterion for measuring learning effectiveness will always be whether the output is trustworthy, and that will not change."
The trust issue is not only about looking at the final result; it also involves a decades-old problem: process data, which in the eyes of fabs and chipmakers is itself core IP and is very difficult to share externally.
"Agents and models are already quite good at understanding textual information, but physical-level modeling remains a huge bottleneck," Gupta said. "Whether small models or large models, they all require massive amounts of data for training. Semiconductor design data is highly siloed, and customers place great value on their own IP assets. Everyone tends to prefer local deployment, but the contradiction has always existed: fabs will not open up core data."
Fabs deliver PDK process development kits to end customers and EDA vendors, but strictly limit access permissions, which creates obstacles for AI model training and data sharing. Coupled with the black-box nature of AI itself, this further raises the threshold for data sharing.
Even so, the scale of agent applications continues to grow. "It can be compared to an engineering team in an enterprise," said Banerjee of Synopsys. "A complete chip or system design will not be completed by a single engineer, but by an engineering team collaborating, managed in layers by supervisors, directors, and executives. The reason for the layered architecture is that each person has their own expertise, and managers coordinate tasks. The goal of early AI in the EDA field was to speed things up and optimize PPA, which is our DSO.ai design space optimization tool; early generative AI was mainly used to generate scripts. For example, if you encounter a parameter problem when using Fusion Compiler and raise a question, the system provides a solution based on a large number of Q&A records, equivalent to a dedicated ChatGPT for EDA scenarios, that is, RTL assistant-type agents."
AI agents can even be used to fine-tune existing steps in the design flow. "We hope to build a set of general-purpose workflow agents," said Wang of ChipAgents. "For example, the design methods for accelerator card chips and CPUs differ greatly, the verification strategies are completely different, and the proportion of formal verification also differs. Flexibility is crucial; workflows cannot be rigid. The platform must support customization for specific chips while also being general-purpose."
Conclusion
"Models carry a large amount of knowledge, and the agents themselves do not need to be extremely smart," said Diamantidis of Cognichip. "Agents can set constraint boundaries and supplement specialized capabilities. But now our thinking has reversed: the hardware field does not have billions of samples, and data is scattered across silos and cannot be centralized. We can train a highly focused model and put a lot of logic on the agent side, but this will make the center of gravity too heavy and the system prone to imbalance. This is one of the core challenges."
"In the future, agents will schedule and run the hardware resources of various agents; the answer is yes," said Faisal of Movellus. "But the process will not be easy. The semiconductor industry, which supports this intelligent system and at the same time implements it in internal toolchains and even resource management, is itself conservative and slow to iterate. The cost of a $500 million tape-out failure is enormous. Various management-type agents will also emerge to manage racks, temperature control, and load scheduling."
Agentic AI will not replace traditional IP, but it will change the process of using IP. Change is coming, and the road ahead remains full of challenges and unknowns.
This article is reproduced from media reports or online platforms and represents the author's personal position or views. Our reproduction is solely for sharing and does not mean that we endorse or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service in a timely manner, and as a neutral platform service provider we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations
Huawei & CNN co-produced documentary "Innovation Unlimited," Episode 1: Small Cells, Great Leap (2003-2008) 2026-11-16 18:03
In October, China's chip output rose sharply by 21.9% 2026-11-16 17:59
Investing $1 billion over 5 years, SK Hynix makes a big move 2026-11-16 17:58