When is the end?

You’ve probably felt the pressure like me: “AI will change everything.” Maybe you’ve even wondered if there’s any point in fighting the current. The promise is that AI will soon solve all the routine tasks for us. But when will it actually happen and what does it really take to get there?

All this hype around AI pushes me to experiment. Not long ago, I wondered where I could actually apply an LLM and I came up with this simple idea: what if I could build an agent that analyzes failed regression tests, saving time on root cause analysis? At first, it sounded promising.

The flow seemed pretty simple:

• Parse the test report (we use Allure, by the way). At this stage, I even thought about ditching it and using something built-in or even vide coding something own.
• Create a prompt, or rather a template.
• Add various types of context to the prompt (I tried everything): test error message, stack trace, error logs from Grafana, known issues from the failed category, description of the endpoint or area (from requirements to technical implementation), etc.
• Send it all to the LLM for judgment.
• As an experiment, post the result to Slack.

But the results didn’t impress me. Adding more context didn’t help much either.

But here’s the main point. My takeaway is this: if you want to build something truly valuable, the AI agent needs to know literally everything that’s going on in the team or teams.

It’s a bit like how a new teammate joins your project. At first, they know nothing. Over time, they pick up enough context to understand what’s happening and can start making sense of problems or at least know where to look. That’s why knowledge sharing and communication are so important in teams.

Same with AI. I used to think, what does it need to know for my use case? But then I realized that for a truly valuable solution AI should always see the whole picture, not just random pieces of context. It should always be up to date with everything: results from all test runs, current manual and automated tests, original requirements, approved requirements, diagrams, architecture docs, hypotheses, discussions, code, PRs, bugs… literally everything.

Sounds crazy. I asked myself, why can’t we just give it everything right now? Technically, you can (like with RAG). But how do you keep all of that updated? Every day, something changes. Manually collecting and refreshing all this is hard, slow and eventually no one will bother.

In my opinion, either there needs to be one platform where everything happens (chats, meetings, code, tests, etc.) or some reliable aggregators (using protocols like MCP or something similar).

And maybe then we’ll get closer to that moment when you really need fewer people. Not superintelligence, but when AI has the full context of your product and how people interact with it.

The debate on whether documentation is needed for your project is already over. AI agents can only use what they understand. Better documentation means better AI performance.

So, when is the end? Maybe it’s not about some final destination. Maybe the real shift is when teams start treating documentation, context, and shared knowledge as their superpower — not just for people, but for every AI agent they hope to unleash. The future belongs to those who keep the lights on for both humans and machines.

Personally, I decided to move on by getting fundamental knowledge in AI, ML and LLM. The hardest part is — I don’t know what to do with it. But when I learn something new, I try to think about how I can use it and do some experiments.