Episode
Agentic Engineering: Generative AI in structural applications
- Published
- Sep 17, 2026
- Duration seconds
- 1319
- Processing state
not_requested- Canonical source
- https://www.designforam.com/p/agentic-engineering-generative-ai
Actions
POST https://stenobird.com/v1/public/podcasts/cdfam-computational-design-symposium-7077587/episodes/agentic-engineering-generative-ai-in-structural-applications/transcription-requests
Idempotently request low-priority transcript generation for this episode.GET https://stenobird.com/podcast/cdfam-computational-design-symposium-7077587/agentic-engineering-generative-ai-in-structural-applications.md
Read the agent-friendly Markdown representation of this episode resource.
Summary
CDFAM Computational Design Symposium — Washington DC 2026 Sergey Pigach · CORE studio | Thornton Tomasetti CORE studio spent a decade building machine learning tools for structural design and analysis, all running as cloud services behind APIs. When MCP arrived, handing those same tools to an agent turned out to be close to trivial. Sergey Pigach demonstrates Bender, an agentic system running on AWS that the firm talks to through Slack: ask it for a concrete column stack and footing for a five-storey residential building in New York, let it make the remaining assumptions, and it calls the tools the engineers use, renders the result and writes a design summary. It lives in Slack deliberately, because an agent sitting in a shared thread already has the context of the conversation around it, which a one-to-one chatbot does not. Specialist sub-agents handle questions like embodied carbon. From there the talk moves to agents talking to each other. A2A is a protocol for delegation between agents, complementary to MCP rather than competing with it, but it has no discovery layer — a public agent the team put online was found by nobody. That gap prompted Waggle, Pigach's own side project, which crawls for valid agent cards and builds a searchable index with health, quality and trust signals, then delegates a request to whichever agent can handle it. He also shows agents paying each other small amounts to cover expensive work. The most uncomfortable result is a benchmark. CORE studio asked its own engineers for the hardest structural problems they could devise, assembled 91 of them, graded answers to within one percent, and gave the models nothing but a calculator and a Python sandbox — no internet, no engineering software. They expected around half. It saturated immediately, wi…