Agentive AI in EDA When AI No Longer Just Assists but Controls EDA Workflows

With the Fuse EDA AI Agent, Siemens EDA aims to reach the next level of AI-driven design automation. At its core is not another copilot but an agentive system designed to plan, execute, and monitor complex workflows across multiple EDA tools.

AI in EDA did not start with generative AI. Machine learning methods have long been embedded in tools to accelerate simulations, make variations more manageable, or optimize individual design steps. What is currently changing is the ambition: AI is no longer meant to assist in isolated areas but to coordinate entire workflows across multiple tools. This direction was evident at several EDA and semiconductor events in 2026. Siemens EDA demonstrates what such an orchestrator could look like with the Fuse EDA AI Agent.

This year, the Fuse EDA AI Agent is a key focus for the company. Amit Gupta, Chief AI Strategy Officer and Senior Vice President and General Manager for Solido Custom IC at Siemens EDA, describes the move toward agent-based AI as a shift on two levels. On the one hand, AI and GPU acceleration are intended to accelerate the compute-intensive core functions of EDA tools, such as simulation, verification, or optimization. On the other hand, agent-based AI is aimed at increasing engineers' productivity by planning and executing complex workflows across multiple tools.

"Generative AI copilots were the industry's first response to this second need, but they are no longer sufficient," says Gupta. Given increasing design complexity and fragmented tool landscapes, manual scripts and isolated point solutions can no longer scale effectively.

The Fuse EDA AI Agent builds on the Fuse EDA AI System and is designed to plan, execute, and adapt workflows in semiconductor and PCB system design across multiple tools. According to Siemens, this spans from early design and verification phases to physical implementation, timing closure, power optimization, DRC analyses, design-for-test, and manufacturing sign-off. With this, Siemens addresses a key bottleneck in many development environments: modern designs are no longer created in a single tool but through long, highly specialized toolchains. The more complex these become, the greater the effort required to coordinate them. Not only do the computations themselves take time, but also the handoffs between tools, the validation of intermediate results, and the feedback loops when a later step reveals issues.

Gupta expresses the ambition as significantly greater than that of a traditional assistance system: "This is precisely what the Fuse EDA AI Agent offers: the transition from AI features within individual tools to autonomous, end-to-end workflow orchestration." This allows customers to shorten design cycles without compromising quality standards. However, the extent of this effect in practice is likely to depend heavily on the specific tool landscape, data foundation, and process maturity within the company.

Why Generic AI isn't enough for EDA

Semiconductor and PCB development is not an environment where a general AI agent can simply be applied to existing tools. The processes are physics-driven, the data formats are specialized, the data volumes are vast, and the IP is highly sensitive. Generic agents quickly reach their limits here. They do not understand the significance of many EDA workflows, cannot reliably configure tools, and may lose context or draw incorrect conclusions when dealing with lengthy toolchains.

Gupta mentions in the interview five hurdles that Siemens had to address in the development of Fuse EDA AI Agent:

• First, chip design is based on physical methods and empirical knowledge, which are hardly reflected in public training data. Therefore, a generic AI agent does not automatically know how to configure an EDA tool or meaningfully sequence a workflow.
• Second, EDA workloads often run on secure local clusters with older schedulers and massive datasets, not in generic cloud frameworks. An agent must therefore integrate into existing HPC and enterprise infrastructures instead of assuming an idealized AI cloud environment.
• Third, the size of modern toolchains can overwhelm standard models. The more tools, parameters, intermediate results, and dependencies a model must consider simultaneously, the greater the risk that the context is no longer processed cleanly or incorrect conclusions are drawn.
• Fourth, many EDA data are in complex, partly binary formats such as waveforms, netlists, LEF/DEF, or GDSII. For AI to extract usable information from them, it requires specialized parsers and clean data preparation.
• Fifth, sensitive IP requires robust security mechanisms. Role-based access, isolated execution environments, audit trails, and approval points by humans are therefore not accessories but prerequisites for productive agentive AI in EDA.

Siemens therefore positions Fuse not as a general AI assistant, but as a domain-specific architecture with an EDA data lake, specialized parsers, RAG framework, role-based access controls, audit trails, and human-in-the-loop control points. The security aspect is particularly important for the target group. Chip designs and PCB systems contain highly sensitive IP. An agent-based system that autonomously invokes tools and processes data must not only be powerful but also remain controllable.

Semiconductor Industry in Europe

Siemens EDA Becomes First Software Partner for Europe's Chip Design Platform

From Tool Invocation to Workflow Orchestration

Technically, the Fuse EDA AI Agent is based on a modular architecture. For this, Siemens breaks down EDA processes into individual sub-processes, which are stored as executable work instructions. These "Agent Skills" describe which steps the agent should execute in which order and what expertise is required. Tools are dynamically recognized and invoked via the Model Context Protocol. A supervisor agent can plan tasks and distribute them to specialized worker agents. The goal is not to automate the entire design process in one step but to initially map clearly defined sub-processes reliably and later assemble them into larger workflows.

Gupta compares this approach to LEGO bricks: once enough of these building blocks in the form of automated sub-processes are in place, they can be combined into more comprehensive multi-tool workflows spanning the entire lifecycle. This marks an important distinction from the marketing narrative of fully autonomous chip design. In its current form, human control remains crucial. Engineers define intentions, monitor results, intervene in critical decisions, and retain responsibility for technical trade-offs.

In everyday practice, this could mean that developers no longer need to manually initiate, configure, and check every intermediate step. Instead, they describe a goal, such as an analysis or optimization, and the system takes over tool selection, workflow planning, execution, validation, and, if necessary, error correction within defined boundaries based on this specification. "What we give engineers back is time and the ability to focus their work on innovation," says Gupta. The operational burden of tool recognition, workflow planning, validation, and troubleshooting is expected to decrease.

The role of engineers would shift as a result. Gupta describes a longer-term vision in which engineers move from task execution to strategic oversight, supervising parallel agents that simultaneously optimize performance, energy consumption, and area. For companies, this raises not only a technical but also an organizational question. Who decides which tasks agents may independently take over? At which points are approvals still required, and how are results transparently documented?

However, some of the preparatory work shifts to the user companies themselves. Agentive systems benefit from clearly defined workflows, consistent data structures, and established approval points. Where processes are highly informal, script-based, or only documented in the minds of individual experts, the implementation is likely to be correspondingly more complex.

Nvidia as an Infrastructure Partner

The Fuse EDA AI Agent supports Nvidia GPUs, Nemotron models, and AI infrastructure from the Nvidia ecosystem. The Nemotron models are designed to work reliably in reasoning and tool-calling while also increasing throughput. This is crucial for industrial EDA applications because an agent does not simply respond once but plans many intermediate steps, invokes tools, and evaluates results. This increases the demands on reliability, computational power, and token costs.

Or, to illustrate more clearly: In a simple chat, you ask a question and receive an answer. With an EDA agent, much more happens, as the agent has to plan, formulate intermediate steps, invoke tools, read back results, evaluate them, correct errors if necessary, and trigger the next step. Each of these steps can involve model calls and thus generate token consumption.

Gupta describes the collaboration with Nvidia as significant. The agent supports Nvidia GPUs, Nemotron models, and NIM infrastructure to execute complex workflows more reliably. Nemotron models offer "high precision and increased throughput for agent-based EDA tool calls." Siemens and Nvidia present the partnership as a step toward durable, autonomous agents that can securely operate complex engineering tools and coordinate tasks across extended workflows.

According to Siemens, Nvidia also uses the Fuse EDA solution in its own chip development. For Siemens, this is an important reference point; however, for users, the key factor remains how well such systems can integrate into existing, heterogeneous EDA landscapes. Especially in environments with multiple EDA vendors, established script landscapes, and strict IP requirements, the introduction of agent-based workflows is likely to be less of a plug-and-play matter and more of a gradual integration process.

Anti-Skepticism

Nvidia Boss Huang Sees AI Business as a Trillion-Dollar Earner

What Becomes Relevant for Users

For semiconductor and PCB development teams, the focus is less on individual product functions and more on the question of how far established EDA processes can actually be automated. Agent-based AI can accelerate routine workflows, tool transitions, analysis chains, and troubleshooting. However, it also highlights where processes are not properly defined, where data is missing, or where responsibilities have so far been informally managed.

This brings data quality, tool integration, access control, and governance into sharper focus. Additionally, the issue of traceability becomes crucial: the more steps an agent autonomously plans and executes, the more important it is to log decisions, tool calls, and intermediate results comprehensively. In productive EDA environments, it is not sufficient for a result to simply appear plausible. It must also be clear and traceable how the result was achieved.

The practical benefit will therefore largely depend on how well companies structure their existing workflows, which data sources can be made accessible, and how precisely responsibilities between humans and agents can be defined. For smaller, well-defined tasks, agent-based AI is likely to deliver tangible benefits more quickly. The vision becomes more challenging, however, when AI agents are tasked not only with automating individual steps but also with weighing multiple design goals against each other simultaneously: performance, power consumption, and chip area (PPA). These factors are closely interconnected; an improvement in one area can create new issues in another.

Gupta sees the transition already underway: "The shift from copilots to autonomous agents is not a distant vision. It is already happening." The Fuse EDA AI Agent should therefore be understood less as an isolated product innovation and more as a signal of a broader transformation. AI in EDA is no longer being conceived merely as an assistive function within individual tools. Instead, it is increasingly intended to connect workflows, prepare decisions, and autonomously coordinate technical processes.

For engineers, the question shifts from "What tasks can AI take off my plate?" to "Which parts of my development process can I entrust to a controlled agent system?" Whether agent-based AI in EDA quickly transitions from a demonstrator to a productive tool will likely be decided precisely at this point: in the balance between automation, control, and trust.