Synopsys and Ansys power the future of innovation—connecting silicon to systems.
EDA
System
IP
SolvNetPlus
Training & Education
Part 1: Hitting the Capacity Wall
AI has already improved productivity across many engineering tasks, helping teams optimize results, accelerate work, and reduce friction in complex workflows.
A different kind of capability is now emerging: agentic systems that can execute engineering workflows against defined goals and constraints.
This is the foundation of autonomous engineering.
Enhance Your Chip Design with AI
Explore the enhanced Synopsys.ai brochure, featuring cutting-edge advancements in Advanced Optimization, Generative AI, and Agentic AI to transform your chip design process.
Shifting from tool use to goal-directed execution
This ongoing evolution has progressed in three broad stages.
First came optimization. In semiconductor design, AI techniques such as reinforcement learning proved valuable in navigating large design spaces and producing better outcomes for power, performance, area, verification coverage, and analog circuit quality. This was a huge breakthrough because it showed that AI could improve engineering results inside production flows. But it was still achieved within the boundaries of a human-managed process. It improved decisions inside the workflow without changing who owned the workflow itself.
Then came assistance. Generative AI brought copilots into engineering environments, making it easier to answer questions, generate scripts, debug runs, and surface documentation or institutional knowledge. These systems reduced friction and lifted individual productivity. They made complex tools easier to use and compressed the time between problem and response. But even then, the underlying model remained the same: the human asked, the system responded, and the human continued to do the work.
Autonomy changes who carries the workflow forward.
With agentic AI systems, the deliverable is no longer an answer or a suggestion. It is the completed workflow. Instead of helping an engineer perform the next step, the system can pursue an outcome across many steps — planning, executing, evaluating, adapting, and continuing until the task reaches an acceptable result.
That changes the unit of work from an individual prompt, command, or recommendation to a goal-directed workflow.
Encoding workflows for machine execution
All of this becomes clearer when we look at what autonomous systems are beginning to do in practice.
Take RTL creation, for example. In a traditional flow, an engineer interprets a specification, writes RTL, develops tests, runs verification, debugs failures, and iterates until the design is functionally sound and ready for the next stage. Each of those steps involves knowledge, tooling, judgment, and repeated cycles of correction.
An agentic system can now take that same objective and execute the entire workflow — from specification to generated RTL, through test creation, failure analysis, debugging, lint cleanup, and synthesizability checks. This involves far more than code generation. It’s a matter of navigating the loop and adjusting along the way — autonomously.
The same principle applies in physical design. Congestion and design rule check (DRC) resolution have historically demanded extensive manual effort because they require more than a static answer. They require diagnosis, interpretation, and iterative remediation. An agentic system can analyze maps and reports, identify likely root causes, create a sequence of actions, implement fixes, and continue refining until the layout is clean. Again, the breakthrough is not isolated intelligence. It is machine-executed workflow logic.
We’re applying this to simulation as well. Many engineering tasks are essentially controlled iteration loops: runàanalyzeàadjustàrerunàcompareàconverge. These loops are structured, but they consume enormous amounts of time because they depend on continuous human facilitation.
Agentic systems can increasingly execute those loops independently, performing many more iterations than a human could reasonably manage. That allows them to explore more options, learn from each run, and converge on better results over time. Those learnings can then be deployed at scale because the knowledge and skills are concentrated in a single system that everyone can use.
Re-engineering engineering
These examples point to a larger change in engineering practice: workflows are being encoded so systems can execute them against defined objectives.
Traditional EDA automation was built around expert-operated tools. Autonomous engineering is centered on expert-defined objectives. The engineer specifies intent, constraints, thresholds, and priorities. The system then determines how to move through the workflow to achieve the goal. In effect, expertise is being translated from something a person applies manually into something a system can operationalize repeatedly.
It’s important to note that this work can’t be performed by a single agent.
Real engineering workflows are too complex to be handled by a standalone, monolithic model. They require specialized capabilities — generation, verification, diagnosis, analysis, planning, correction — that come together in a coordinated fashion. Tomorrow’s intelligent products will be designed, tested, and verified by systems of purpose-built AI agents, each contributing to the progression from intent to outcome.
None of this removes the need for human expertise. In fact, it raises the value of expertise.
Human engineers will still define the objective, establish what “good” looks like, evaluate tradeoffs, and decide when an outcome is truly acceptable. And as workflows are re-engineered so they can be executed autonomously, humans shift from doing the work to orchestrating the work at higher, more impactful levels.
And once that becomes possible, the key questions become: How can agentic systems be assembled end-to-end for the greatest engineering speed, efficiency, and outcomes? And what expertise — spanning humans and machines — will be required to deliver the greatest strategic and operational advantage?
Part 1: Hitting the Capacity Wall
Part 3: Why Knowledge Becomes the Advantage (coming soon)
Part 4: From Execution to Orchestration (coming soon)
Continue Reading
