I'm curious about how people are actually learning to create real-world AI agents that integrate into business workflows, manage finances, handle contracts, and take on real responsibilities. I'm not talking about simple chatbots or hobby projects; I mean serious agents that can operate under legal constraints and accountability.
Here's what I'm grappling with: there seem to be very few reference repositories for serious AI agents, and much of the existing content is either too shallow, scattered, or only focuses on orchestration. Blogs mention agents but often skip over critical aspects like accountability, rollback, audit, or how to handle failures. Most genuine systems appear to be proprietary and not accessible to the public.
I get that developing these systems can be risky, which is probably why they're not frequently open-sourced. But it's clear people are building them, and I want to learn from those with direct experience.
How did you get into this field? What should newcomers focus on studying first, whether that's infrastructure, systems design, or understanding the legal landscape? Are many teams basically just embedding LLMs in traditional software? How do they manage responsibility, human oversight, and failure in production? And where can I find serious discussions about these topics? I'm not looking for easy shortcuts or magic solutions but want to develop a solid mental model and structured learning path for creating production-grade systems.
4 Answers
If you want to grasp how AI tools work, I suggest avoiding standard frameworks like Langchain initially. Instead, start with the OpenAI API and build everything manually. Frameworks can complicate things for newcomers since it can be hard to distinguish between what's AI and what's part of the framework. Think of it like Kung Fu Panda: there’s no secret ingredient; it’s just LLM + tools, orchestrated with basic coding logic. For legal reviews or sanity checks, integrate different AI models to cross-verify outputs. But be prepared—this can require a lot of resources, including time and money, and might involve outsourcing tasks to teams in low-cost areas. The missing link in creating efficient AI tools seems to be a standardized API structure, like MCP, which could make it easier to integrate various tools without reinventing the wheel.
AI often seems like a solution in search of a problem. Most companies that try to implement AI agents end up abandoning the project because it just doesn’t deliver the anticipated results.
From my experience in a major tech company, I've created a few tools to help developers work with APIs using LLMs and other agent tools. However, I've noticed that while people acknowledge these tools, they don’t often want to actually build the integrations. It feels like a lot of fluff right now because, honestly, the current AI implementations haven't taken off in DevOps; there's just not enough that they can do that can't be handled by existing automation tools. It's less reliable than automation, and it's often more work to implement.
Thanks for the insight! It's exactly this kind of perspective from real DevOps experience that I was hoping to hear.
Interesting take on the whole AI narrative!
Same here! I work in a similar tech environment, and it’s like people say, 'Yeah, I built the server,' but they keep falling back on the standard CLI tools instead.