The most tedious part of 3D printing isn't the printing itself — it's the endless back-and-forth between design and manufacturability. Engineers spend hours fixing overhangs, thin walls, and poor layer adhesion. Now, researchers at Carnegie Mellon University have built an AI that does it for you.<br><br>Meet AgentsCAD, a multi-agent system that combines geometric reasoning with large language models to automate Design for Manufacturing (DFM) modifications for FDM parts. It parses a STEP file to extract geometry, then uses a pipeline of LLMs to detect issues and suggest fixes. Claude Sonnet recommends reorientations, fillets, or chamfers, while GPT-4o checks that the repaired geometry still makes sense. The system outputs a modified STEP file and a human-readable report.<br><br>This isn't just another CAD plugin. According to the preprint posted July 2, 2026, AgentsCAD is the first system to close the loop between defect detection and automated geometry repair. It detects unsupported overhangs above 45° using face-adjacency graphs, and optionally labels features with a GraphSAGE model trained on 59,665 parts. That's a lot of data — and a clear sign that AI is getting serious about physical design.<br><br>For the average maker or engineer, this could mean fewer failed prints and more time on creative work. For the industry, it's a glimpse of a future where design and manufacturing speak the same language — without human translators. The system currently focuses on FDM, but the approach could be adapted for stereolithography, laser sintering, or even metal additive manufacturing. Trends suggest we'll see similar multi-agent DFM logic infused into platforms like FreeCAD or Onshape within a couple of years.<br><br>But don't throw away your calipers just yet. This is a preprint — not peer-reviewed. The technical community will need to validate how well AgentsCAD handles organic shapes, large assemblies, or safety-critical parts. LLM-based geometric reasoning can hallucinate: it might generate an elegant fix that is, in fact, unprintable. The paper doesn't mention an open-source release or benchmarks against tools like Autodesk Netfabb. Expect follow-up studies that compare repair quality and print success rates.<br><br>The big winner here? Anyone who hates clicking through CAD menus for the same fix over and over. The losers? Legacy DFM consultants — and maybe the ego of engineers who think AI can't understand geometry. AgentsCAD won't replace you yet, but it could become your most opinionated intern. The question is: when it suggests a modification, will you trust the algorithm — or your gut?