Flexcompute Introduces Autonomous Photonic Design Loop Revolutionizing Chip Engineering

Flexcompute Unveils Revolutionary Agent-Driven Loop for Autonomous Photonic Design

Flexcompute, a pioneering physics company, has made a remarkable breakthrough in photonic chip design by introducing the first fully autonomous agent-driven loop for end-to-end design processes. This innovative approach allows AI agents to propose designs, conduct physics simulations, verify fabrication constraints, and iterate the design without requiring human intervention at each step. As a result, the technology delivers tapeout-ready layouts in a matter of hours rather than the weeks traditionally required.

Photonics is widely recognized as one of the most complex fields in multi-physics engineering, particularly when it comes to co-packaged optics for AI data centers. This endeavor requires the integration of various physical principles including optical, electrical, and thermal physics across numerous configurations. Conventional simulation cycles often take days to complete for each iteration. However, Flexcompute's new solution, PhotonForge, directly connects agents with foundry process design kits, layout options, and circuit simulations, thereby bridging the gap between device physics and system-level performance. The Tidy3D solver, which is GPU-native, significantly accelerates this process, enabling rapid closure of the design loop.

Vera Yang, the President and Co-Founder of Flexcompute, stated, "We did not simply link our tools to a large language model (LLM); rather, we constructed a framework that empowers reasoning agents to access physics simulations and fabrication checks directly. They can autonomously manage the entire design process based on challenges posed by our engineering team." The designs generated by these agents maintain expert engineer quality and are produced with minimal human involvement, substantially increasing efficiency in the design process.

The architecture supporting this revolutionary system comprises several key components:
1. Tidy3D Multiphysics Solver - Compresses days into minutes for each design point.
2. Python APIs and Flexagent MCP Plugin - Provide agents with immediate access to documentation and refined workflows.
3. Dedicated GPU Cluster - Eliminates queuing delays, allowing agents to run multiple simulations simultaneously without needing human input.
4. PhotonForge - Supports over eight foundry Process Design Kits (PDKs), ensuring every design is production-ready from the first iteration.
5. State-of-the-art LLMs - Enhance the educational capabilities of agents regarding coding and physics nuances, enabling them to manage the design loop effectively.
6. Curated Example Library - Offers a range of validated reference designs for the agents to learn from and apply in new projects.

Recent demonstrations of the system have included a variety of devices, from passive components to active modulators and RF components, as well as full layout generation. The end-to-end multiphysics simulations have allowed agents to resolve challenges that previously required senior photonic engineers, achieving completion in a single overnight run.

Flexcompute is one of the first organizations to utilize autonomous AI engineers powered by NVIDIA NemoClaw, a recent announcement made at COMPUTEX 2026. This innovative workflow is compatible across various platforms including Anthropic's Claude and OpenAI's Codex, enabling clients to choose their preferred agent without reconstructing the underlying engineering framework.

Looking ahead, Flexcompute’s roadmap entails a progression from simulating individual chips to co-designing comprehensive systems. The current phase, from 2016 to 2023, has successfully delivered GPU-native, multi-scale, and multi-physics simulations. The ongoing phase aims to integrate design across PDKs, simulation, layout, and verification within a singular canvas-based framework. The forthcoming phase is set to enable AI agents to generate entire chips based on human intent and specifications, marking a substantial advancement in the industry.

Additionally, this architecture can be applied across the entire physics stack of Flexcompute, introducing agent-driven design across various domains including electromagnetics, heat transfer, charge transport, and computational fluid dynamics.

In conclusion, Flexcompute is not just redefining photonic chip design; it is heralding a new era in the way complex engineering problems can be approached through artificial intelligence.