--- title: "Why I Build Platforms, and Why Agents Need One Now" canonical: "https://www.guild.ai/blog/community/why-i-build-platforms-and-why-agents-need-one-now" --- # Why I Build Platforms, and Why Agents Need One Now - **Category:** Community - **Author:** Vincent Durmont - **Published:** Nov 11, 2026 - **Reading time:** 45 min A few years ago, I was at a fintech startup, watching a customer do something that got my attention. They were one of the earliest adopters of Devin, and they were running huge code migrations in record time. Work that would normally take a team months was getting done faster than anyone expected. That stayed with me. I started experimenting with coding agents, and a picture began to form of what would be possible with autonomous agents that take care of the work you don’t want to do and don’t want to hire for. ## The platform problem Most of my career has been spent building platforms and developer tooling: internal tools at Facebook, blockchain infrastructure at Novi, payments infrastructure at Lightspark. I’m drawn to products that serve as a foundation for many different uses, where the value comes from what other people build on top of them. When we started Guild, Claude Code had just come out, and we were at the very beginning of the AI explosion. There was no good way to build autonomous agents and run them the way you’d run production software. That gap is what pulled me in. Building the platform where agents live and operate like real production systems was exactly the kind of problem I wanted to work on. Some things have played out the way we expected. Models have improved incredibly fast, but they still need a solid harness to govern them. We still see agents doing things they shouldn’t, and guardrails aren’t optional. We also expected that an agent isn’t always the right tool. Agents are non-deterministic, and some parts of a company’s processes need determinism. That’s why we built Guild so that code agents (written in TypeScript) work seamlessly alongside LLM agents. Combining the two lets us improve the speed, cost, and quality of a workflow all at once, rather than trading one off against the others. What has surprised me on this journey so far is that a single agent might not be the right unit at all. What actually solves people’s problems is often a suite of agents, plus triggers, plus context, plus views. ## From “human in the loop” to production Most AI work today is still experimentation, and experimentation usually means a human is watching. Someone types a prompt into Claude Code or Codex and follows the agent as it works through the problem. They can stop it, answer its questions, approve, or deny its tool calls. Deploying autonomous agents in production is a different thing. These agents run without supervision, so you can’t count on a human to catch problems. You have to treat them the way we’ve always treated production software: identity, testing (evals), optimization, governance, credentials management, monitoring, observability, escalation. None of these concepts are new. Software engineers have worked with them for decades. They’re still valid; they just need infrastructure that fits what agents can do. That’s what we’re building at Guild. ## What matters in engineering now My engineering philosophy has changed a lot. The details of code (formatting, style, and so on) basically don’t matter anymore. AI writes the code. What I still care about deeply comes down to two things: how you store your data and what API you expose. If you can clearly articulate what each piece is, what role it plays, and how to interact with it, the business logic is the easy part. With AI, it’s practically free. Those two things matter because they’re expensive to get wrong. Data that wasn’t designed well is hard to fetch efficiently, and no amount of generated code fixes that. Code is cheap to change; data is expensive and painful to migrate. And an API is a contract with your customers. You can’t change it casually, so you’d better be clear about what you’re promising. I also watch closely for what I call the decoherence trap: the slow buildup of models inside a system that are each valid on their own but incompatible or redundant with each other. AI models are good at solving local goals and bad at seeing the whole system. Left unchecked, they create pockets of conflicting or duplicated logic that degrade the customer experience over time and get harder and harder to untangle. Engineers have to set the guardrails and really understand what they’re building to prevent that. ## Every engineer is a senior engineer now AI levels up the job. What matters is systems, architecture, deciding what to build, and setting direction. In a sense, every engineer is now a senior engineer. If you’re entering the field today, I’d also invest in product skills. Deciding what to build and how to test it with real customers is becoming more important as the bottleneck shifts from building to selling. Early in my career I believed the old saying,  “If you build it, they will come.” My thinking now is that most startups spend too much time on their own view of the world and take too long to listen to their customers. At Guild, we have a clear vision of where things are going, and our platform reflects it. Balancing the future you’re building toward with what customers need today is the hard part. ## The best agent I’ve seen Guild’s self-healing factory sends production logs to a storage provider. Agents analyze what’s happening, deduplicate issues, and maintain the backlog. Other agents pick issues off that backlog and fix them. A large part of Guild now fixes itself. And because everything on the platform is customizable, we tuned it to fit our exact processes and conventions. That’s the point I care most about. Getting great results from AI isn’t magic. It takes tuning. Take code review: ask a generic agent to review your code and you’ll get a lukewarm job. But when the agent is customized for what your company cares about, the details of your system, the verbosity you like, and the workflow that works for your team, it’s a completely different experience. We’ve seen it ourselves. Guild gives you strong agents out of the box. The real strength is the ability to make them better, both manually and automatically, for your company, your team, and your workflow. ## If not this? If I wasn’t currently leading engineering at Guild, I’d still be building software. I love building, and I love solving problems, whether that’s in AI, finance, crypto, social media, or anything else. Though lately I’ve gotten pretty into space exploration. Maybe in another life. 🚀 ## Download the 2026 Agent Management Gap Report Find out what technology leaders had to say about their own agent platform problems and why enterprises are scaling AI agents faster than they can see, secure, and control them Get the guide