Building ToyotaGPT: 50+ Production Agents, One Config File, Zero Architecture Reviews

Hi, everyone. I'm the head of Agentec AI and product research at Toyota. I have Kotal joining, who's the head of AI engineering. So in 2023, when generative AI just changed the world, and inside Toyota, 65,000 people, multiple factories, and one question that we have is, what do we do with this? We are the Enterprise AI team. Every AI use case that's going to production comes to our team. So one standard, one platform, no chaos, no duplication. The problem is nothing existed before that could be what we wanted to do. So we opened up a blank repository in our GitHub and built it ourselves. So this is the story today I'm going to talk about. I'll show you how we built Toyota GPT, a platform that ships AI agent in very few days. Then Kotal talks about connects with our Toyota production system and how LangChain, LangGraph, and LangSmith is really useful today in our ecosystem. That part really surprised you. Let's go. So in 2023, every team at Toyota are rushing to build their own chatbot. Same ingestion, same extraction, same pipeline. Multiple duplicate things are coming up. So no security standards, no architecture standards. Pure duplication at an enterprise scale. So what we stopped in, and our job is to stop that. So be the platform every AI agent built. The pressure was real. The timeline was yesterday. So what we need to do? We need to go for work. So one rag chatbot at Toyota used to mean six engineers, six months in the early days. Not because AI was hard, because everything around it is hard. Security reviews, architecture sign-offs, and injection plumbing across the multiple data sources and rebuild from scratch. So our delivery was stuck in months. So that gap, we decided to close it permanently. So four days, we built a dynamic graph creation approach on LangGraph, give it the use case, give it the data connectors. The entire graph builds it automatically. So React agents, deep agents, everything plugged in. So no security review, no architecture review. It's all reviewed at once because the architecture never changes, the security never changes. The only difference every AI agent we built is just a config file. So six months became four days, six engineers became one. That is not an optimization. That's a different way of building. So the thing that kills every AI agent before it even starts is the extraction. The bad text in, the bad text out. Our data at Toyota is brutal. PDFs, words, Excel, CAD files. AutoCAD, you name it, we have it. Scan manuals from 19s, Toyota, tables inside tables, inside tables, inside tables, inside images. We have complex data sources that we have in Toyota because we have Japanese, we have English, we have a lot of different languages. You name the format, we built an extractor for it. So layout, aware parsing, and OCR with vision. Schema mapping, every source, one unified index. And we built it ourselves, cutting enterprise license costs. So LangGraph at the core, thanks to Harrison Chase for building the great framework. LangSmith for observability. Vector databases, SharePoint, and DM&A cybersecurity team baked in from day one working with us. Every agent inherits this and exposes via API. Toyota GPT web, internal apps, factory machines, robotics. which uses LangGraph in the background, vision language action models, and everything runs on top of this framework. The pipeline is LangGraph and end-to-end routes dynamically by source type. But the concept I want you to take home is skills. A skill is a unit of intelligence. We do two things nobody else does at scale. One, enterprise-gate skills shared across every agent. One library, no drift. no duplication and two we generate skills automatically from unstructured data we have terabytes of data sitting in our databases feeding these documents the pipeline synthesis skills emerge without a single engineer writing them by hand and on top a unified tool layer mcp compatible every tool secured for any ai agent you want it it's there it's secure already done So today, over 50-plus agents are in production, every single build on Toyota GPT. Every one is a config file. From battery brain on the Plan 4, and Tukura, the long-term memory of the entire enterprise, a Toyota way, Toyota's culture and principles codified into an agent, and two-vehicle AI expert that tells about every Toyota model, every spec, every history, and all queryable in seconds. Let me walk you through a few to show you this platform. actually makes most possible. GearPol. GearPol started as a hackathon idea. One of our team members, Braden Bufford, had a vision. What if every manufacturing engineer could just type their problem and get the answer instantly? Today, GearPol sits on terabytes of data in our vector databases, serving every manufacturing plant across North America. His reality is replaced. When a production line engineer goes down, an engineer walks to a bookshelf and pulls a manual, flips through the pages, manually searching all the information, and fix the problem. That takes hours and sometimes days. But if a production line stops for a few hours, we lose millions of dollars because we are not making cars. So today, type the problem, get the solution in 10 seconds. From a hackathon idea to millions of dollars savings. R&D GPT. This is close to my heart. Every color you see in Toyota on the road, we created that from scratch. We created the paint from scratch. That years of R&D testing in extreme colds, extreme heat, to make sure the quality doesn't degrade. A year, sometimes two, sometimes two to four years. Today, R&D GPT learns from decades of our own past research. Deep research is really helping us to solve that. Now techniques emerging from old technology, old knowledge, what we take multiple years, no compression because of our own institutional knowledge is now searchable, connectable and queryable within seconds. And then Katya GPT, the design agent that lives inside the tool designer, already uses new car designs, existing parts, existing designs, can query best practices, find references, identify patterns, without even leaving their canvas. So, zero context switch. The AI is just there in the workflow exactly where it needs to be. When we started our journey with Harrison, nothing existed with us. We built the entire ingestion, we format from PDFs, from Kedia files, the Langshan Reli ecosystem. came into the life, and then we built the framework using LangChain, LangGraph, and DeepAgents, dynamic graphs, automatically populating. We built the skills engine, intelligence that generates itself. We built a tool layer, MCP ready, enterprise secured. We took a hackathon idea and turned it into a terabytes of production. 50 agents, multiple dollars of savings, built from zero, from scratch. So we learned every single thing that's happened. wait for the industry to catch up, we went ahead and built it. So now, Kotal is going to show us something genuinely surprises everything we built. Toyota invented the philosophy behind it, I think, factory floor in 1988. Kotal, it's yours. Thanks, Ravi. Toyota is arguably the best automotive manufacturer in the world, and by extension, one of the largest and best hardware manufacturers. And it got there through something called the Toyota Production System. The Toyota Production System or TPS is a philosophy, a framework for building a lot of anything really, really quickly and with really minimal resources. So by extension, it's a philosophy on how we can build vehicles in a manufacturing line with very leanly, with minimal staffing, minimal resources, make a manufacturing line modular, make it robust so that it's amenable to breakdowns. and keep continuous flow from raw materials all the way out to when a car comes out on the other end of the manufacturing line. The principles for TPS are really the backbone for any scaled hardware manufacturing line that we see today. It started with Toyota. It's been fashioned over the course of almost 100 years, but really became formalized in the 80s, as Ravi mentioned, in order to help Japan, who had a lot fewer resources in North America, compete in a North American automotive front. And TPS is kind of the backbone for all hardware manufacturing today. And we see a very similar manner on TPS with LaneChain. LaneChain is the modern backbone for which all next generation software and agentic workflows will be manufactured. The principles from hardware manufacturing translate really well into agent manufacturing. It's just a matter of the substrate or the matter that we're manufacturing. And so this has been a really pleasing and quite an awesome experience, not only to use LaneChain's products, but to become more embedded in their ecosystem because they embody the ethos from which Toyota was founded and all of the principles that our team shares that we work with every day. So a couple of principles of TPS that are pretty obvious to identify with LaneChain is the Andon board. The Andon board in manufacturing is a way to see what's what's going on really quickly without having to survey the whole manufacturing floor what's broken down what's what needs supplies uh what's going well where can i allocate resources in order to bolster another part of the manufacturing line that's dwindling and we langsmith is the literal embodiment of an and onboard. We can see observability over all of our agents in real time, understand what tool calls are working, what features should we focus on for the next PR, for the next product release, and what's going well with our users. What are the frustration points? How do we better serve our users and appreciate our software? Langsmith is the direct analog for the and onboard. One term you all are probably familiar with is the term of Kaizen, which is continuous improvement through slow and steady, but consistent, consistent modifications. And the software engineering culture really embodies that. We're always pushing PRs, we're always bolstering the new products, and we push updates regularly, sometimes nightly. Great thing about Kaizen is that from a Langsmith perspective or a Langchain perspective, there's really a macro level that Kaizen is being implemented and a micro level. From a macro level, the software is always improving. Harrison just announced a bunch of new features today that are going to be a huge advantage to the ecosystem that's already been published. But then from a micro level, there's agents that are continuously improving A very rudimentary example might be the React agent, something that's always monitoring its output and continuously improving it to make sure before it presents to the user the final response, it's actually correct and it does what the user intended. So this philosophy of continuous improvement through steady changes and consistent changes is something we embody with Toyota and we're delighted to see with lane chain throughout the whole ecosystem. My personal favorite is the principle of Judoka. uh translated literally it means automation with a human touch and what lane graph does really well is it automates a lot or abstracts a lot of the uh nuances and monotony that as an engineer i don't want to have to deal with or i don't i don't care to deal with but it gives me it keeps me in the loop it keeps me plugged in so that i still have value as a human and i can still guide the product as it's being developed and deployed and Judoka is really like a handshake deal between AI, automation, and a human to say, I understand each role that you play and I understand we're going to have to adapt as technology progresses. On a manufacturing line, Judoka means that a human understands we need automation in order to manufacture things very leanly and very efficiently. But as technology evolves, a human's role will change because that automation will change. But a human is still critical to ensuring high quality of products and delivery to the final customer so uh yeah lane graph is a literal embodiment of of the judoka principle which we love the next term is uh genchi kombutsu so uh this means to literally go to the source and understand what's going on try to figure out the the root cause of the problem Like, don't just, you know, we can't figure things out on a team's call. If there's a manufacturing issue in Texas, we can't sit in California and try to figure out what's going on. The best way to solve the problem is to go to the manufacturing line, actually touch the hardware, understand the root cause of the problem, and then proliferate the solution throughout the rest of the manufacturing line so that we Kaizen the remedy. Langsmith traces are a direct embodiment of Genchi Kabusu. For every query that I have, every tool call, I can see the entire trace, the entire route to the solution. And if there's an issue, I can see exactly what caused it. And so it breaks out from the Andon board by giving me a direct insight into what the problem is for any one of my products, my bots, my agents, and helps me as an engineer not have to sift through logs or spend a lot of time debugging. I can go directly to the problem, try to solve it. and keep our products up and running for our customers. LaneChain is a direct embodiment of the Toyota production system philosophy. And while the framework is timeless and TPS is the bedrock for all hardware manufacturing today, it hasn't been well translated into software or the manufacturing of agents, but until LaneChain. LaneChain carries the legacy of TPS forward to allow all these principles to be carried over into manufacturing of the new era of software engineering. And so I'm really excited to see where Langchain is going to go because our cultures are so similar. And I fully believe that in the next few years, and actually probably now, the entire AI industry will look to Langchain as the TPS or the Langchain production system, the bedrock from which all these SaaS services, companies, et cetera, the entire industry is built on.

Ravi Chandu Ummadisetti and Kordel France from Toyota's enterprise AI team explain how they built ToyotaGPT — a platform that reduced agent delivery from six months and six engineers to four days and one engineer, with over 50 agents now in production. Ravi walks through the architecture: dynamic LangGraph-based graph generation, auto-synthesized skills from unstructured data, and an MCP-compatible tool layer secured by TMNA's cybersecurity team from day one. Kordel then draws a striking parallel between TPS — Toyota's 100-year-old manufacturing philosophy — and LangChain's ecosystem, mapping Andon boards to LangSmith, Jidoka to LangGraph, and Genchi Gembutsu to trace-level debugging.

Chapters:
0:00 Introduction and the 2023 AI chaos inside Toyota
1:40 Why one RAG app used to take 6 engineers and 6 months
2:09 Dynamic graph creation on LangGraph: 6 months becomes 4 days
2:47 The extraction problem: Toyota's brutal data reality
3:38 The ToyotaGPT stack: LangGraph, LangSmith, MCP, and TMNA security
4:18 Skills: units of intelligence generated from unstructured data
5:06 50+ agents in production — every one is a config file
5:39 GearPull: from hackathon idea to millions in manufacturing savings
6:37 R&D GPT: compressing years of paint research into seconds
7:25 KadyaGPT: AI embedded inside the designer's canvas
8:58 Kordel France: TPS meets LangChain
10:14 Andon board = LangSmith observability
11:47 Kaizen = continuous improvement at macro and micro scale
13:04 Jidoka = LangGraph's human-in-the-loop design
14:11 Genchi Gembutsu = LangSmith trace-level debugging
15:21 LangChain as the TPS for the AI era

Resources:
→ LangGraph: https://www.langchain.com/langgraph
→ LangSmith: https://www.langchain.com/langsmith
→ LangChain Academy: https://academy.langchain.com