At Converge 2026, Synopsys unveils AgentEngineer, a multi-agent system that automates SoC verification and design optimization.
What AgentEngineer Is Aimed At
At Converge 2026, Synopsys unveiled AgentEngineer, a multi-agent system built to automate two of the most labor-heavy phases of chip development: SoC verification and design optimization. In electronic design automation (EDA), those steps sit between a finished RTL or netlist and a tapeout-ready implementation. Verification asks whether the design meets its functional and performance intent under a wide range of conditions. Optimization asks how to improve power, performance, area, and timing without breaking that intent. Both traditionally depend on expert judgment, long tool runs, and many manual handoffs between scripts, reports, and review meetings.
A multi-agent approach splits that work into specialized roles—planning test coverage, driving simulations or formal checks, interpreting failures, proposing fixes, and re-evaluating the design after each change—so agents can coordinate rather than force a single monolithic flow. The “5x leap in EDA productivity” framing in the launch points to reduced human thrash: less time spent stitching tools together and more time spent on decisions that still need an engineer’s judgment.
Why Multi-Agent Automation Fits SoC Verification
SoC verification is hard because the state space is huge and the failure modes are diverse. A block that passes unit tests can still break when integrated with memory controllers, interconnects, security islands, or software-driven boot sequences. Engineers spend large portions of a project writing testbenches, ranking coverage holes, triaging regressions, and deciding which bugs block the schedule.
Agent systems can help by owning the repetitive loop: generate or select stimuli, run the appropriate checks, cluster similar failures, and surface the highest-signal issues first. That does not remove the need for golden specs, clear interfaces, or human sign-off. It does change the default workflow from “engineer drives every step” to “engineer reviews agent proposals and steers priorities.” Teams that already treat verification as a pipeline—stimulus, scoreboarding, coverage, debug—will map more cleanly onto agent roles than teams that still rely on one-off scripts with little structure.
Design Optimization as a Closed Loop
Design optimization is a search problem under constraints. You can improve timing by restructuring logic, adjusting floorplan hints, or changing clocking—but each move can hurt power, area, or routability. Manual optimization often looks like a series of local edits guided by reports that are already stale by the time a human finishes reading them.
An agent-driven loop can treat optimization as continuous experiment management: propose a change set, evaluate it with the same metrics the team already trusts, keep improvements that pass quality gates, and roll back changes that regress critical paths or power budgets. The value is not in inventing new physics; it is in shortening the propose–measure–decide cycle and keeping a consistent record of what was tried. That record matters when later silicon bring-up or a process corner shift forces you to revisit earlier tradeoffs.
- Keep human ownership of architecture decisions, safety requirements, and tapeout criteria.
- Feed agents stable metrics (coverage goals, timing endpoints, power envelopes) rather than ad hoc spreadsheet targets.
- Require every agent-suggested change to be reproducible in your existing EDA environment.
- Review failure clusters and optimization diffs with the same rigor you apply to peer code review.
How Teams Should Evaluate a System Like This
Productivity claims only matter if they show up on your critical path. When assessing AgentEngineer—or any multi-agent EDA assistant—map it to concrete workflows you already run: regression triage, coverage closure, ECO-style optimization, and sign-off readiness. Ask how agents hand off state between steps, how results are audited, and how a human can interrupt or reverse a sequence when a proposal looks wrong.
Start with a bounded pilot: one subsystem, a fixed metric set, and a clear comparison against your current process length and defect escape rate. Preserve your golden flows as the source of truth so agent automation becomes an accelerator on top of trusted tools, not a parallel unverified path. Used that way, a multi-agent system for SoC verification and design optimization can shrink the busywork around EDA while leaving final accountability where it belongs—with the design and verification teams shipping the chip.