, /PRNewswire/ — Flexcompute, the physics company, today introduced the first practical fully autonomous agent-driven loop for end-to-end photonic chip design. AI agents now propose designs, run physics simulations, verify fabrication constraints, and iterate without human input between steps, producing tapeout-ready layouts in hours instead of weeks.
Photonics is one of the hardest multi-physics domains in agentic engineering. Co-packaged optics for AI data centers requires coupled optical, electrical, and thermal physics across thousands of configurations, and conventional simulation cycles take days per iteration. PhotonForge gives the agent direct access to foundry process design kits, layout, and circuit simulation, bridging the gap from device physics to system-level performance. Tidy3D supplies the GPU-native solvers fast enough to close the loop. The result is the first autonomous design environment where every iteration is fab-aware from the start.
“We did not just connect our tools to an LLM. We built a system that gives reasoning agents direct access to physics simulation and fabrication constraint checking, and let them run the full design process autonomously on problems crafted by our engineering team,” said Vera Yang, President and Co-Founder, Flexcompute. “The designs the agents produced are expert engineer quality, delivered in hours, with minimal human involvement.”
Six pieces made this possible. The Tidy3D Multiphysics solver compresses days into minutes per design point. Python APIs and the Flexagent MCP plugin give agents inline access to documentation and curated workflows. A dedicated GPU cluster removes queuing friction, so agents launch dozens of simulations per iteration without human-in-the-loop. PhotonForge support for 8+ foundry Process Design Kits (PDKs) across silicon, silicon nitride, and thin-film lithium niobate keeps every design fab-aware from iteration one. State-of-the-art LLMs from Anthropic, OpenAI, and others now reason well enough about code and physics to drive the loop. And a curated example library of vetted reference designs gives the agent ground truth to learn from.
Recent demonstrations span passive devices, active modulators, RF components, photonic circuits, and full layout generation, with end-to-end multiphysics simulation feeding every iteration. Agents solved problems that previously required senior photonic engineers, completing them in a single overnight run.
Flexcompute is among the first companies to deploy autonomous AI engineers with NVIDIA NemoClaw, announced today at COMPUTEX 2026. The same agentic workflow runs across Anthropic’s Claude, OpenAI’s Codex, and other frameworks, letting customers standardize on their preferred agent without rebuilding the engineering stack underneath.
The roadmap moves from simulating chips to co-designing them. The first phase (2016 to 2023) delivered GPU-native multi-scale, multi-physics simulation. The second phase, in production today, unifies design across PDKs, simulation, layout, and verification on a single schema-first canvas. The next phase, beginning now, turns that canvas into a co-designer, with AI agents producing entire chips from human specifications and intent.
This same architecture extends across Flexcompute’s full physics stack, bringing agent-driven design to electromagnetics, heat transfer, charge transport, and computational fluid dynamics.
Media Contact
Steve Defibaugh: [email protected]
SOURCE FlexCompute Inc.
