← sources

Field report: agentic coding in an enterprise project

source: reddit (reddit:ClaudeAI)published 2026-11-11 10:38ingested 2026-11-12 02:37

retained text

**TL;DR:** We tried a fully agentic coding setup on a real enterprise project. Four weeks, lots of API costs, and one 5x performance regression later, we stopped it.

Most of what I read here is about small or internal projects, so I wanted to share a different perspective.

I work at an IT service provider with about 10k employees. We build software for insurance companies, banks, carmakers and many more. Everyone here knows AI is changing everything, and nobody wants to fall behind. So the company funded an internal agentic coding tool: 100k budget, a team of 20 experienced people.

The idea is simple. You throw a message, a Jira ticket or a full spec at it, and a chain of agentic loops takes over: planning, architecture decisions, implementation, testing. Each loop writes a report for the next one, everything feeds into a shared project knowledge base, and an inbox pings you when something needs your approval. It runs on Anthropic models via API.

Setting it up for our project took a while, since it’s a big codebase with a lot of features. But we got it running, and our first test ticket was: *center a button.*

Ten minutes later we had a merge request. It also had a 5k-line diff and had cost $5 in API fees. A human would have needed one minute.

Fine, bad test. So we tried something real: write an e2e test for a new feature (the project already had plenty, so no new patterns needed). About an hour later, a giant .md file landed in my inbox. Already the first paragraph was full of deliberation about things that didn’t matter. I was tired, I didn’t read it, I clicked approve. Thirty minutes later the test worked. It also cost over $45.

We kept going for a few weeks. Then today we stopped.

First, we looked into the knowledge base. It was full of rules that contradicted each other, all written by the agents themselves. My favorite: an agent refused to touch an “accept terms of service” button, because it didn’t know the terms of service and therefore wasn’t “allowed” to click it. We had a lot of moments like that where we laughed out loud.

The real dealbreaker was performance. Our software generates documents via an API, and in a load test, producing 10k documents took 5x longer. Finding the cause took forever, because the agent had written functions thousands of lines long that no human wanted to read.

**What I took from it**

An old enterprise project isn’t a clean slate. It has years of tech debt and architecture patterns that contradict each other, and a lot of it isn’t logical, it’s historical. I need to metion here: other teams had much better results on greenfield projects.

And then you’re stuck in a dilemma. If nobody looks at the code directly anymore, you need strong guardrails, because a data protection mistake in a product sold to a big insurer is no joke. But the more guardrails you add, the slower and more context-heavy the agents get. They start writing pages of deliberation to the “human in the loop”, so much that the humans can’t keep up. And you can’t remove the human, because someone has to be accountable in the end.

Maybe our approach was just not good. If you have a better one, I’d love to hear it.

To be clear: a lot is changing. Nobody at my company writes code by hand anymore. We’ll need far fewer people, especially far fewer pure coders, and I think roles will merge, with humans becoming more like observers. Nobody knows what things look like in 10 or 20 years. But code is still read. It’s still working *with* the AI, not the AI working for me.

So when I read that software engineering is solved, that everyone can build anything now, that every IT company is dead because you can just build your own Photoshop: it’s not that simple. Software engineering is so much more than writing code, and an enterprise project is not a cool demo.

I’d really like to hear your experiences.

document fetched 2026-11-12 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xteklh/field_report_agentic_coding_in_an_enterprise