How I use Claude Code, MCP and autonomous agents in daily development and automation: specs before code, context control, approval gates, monitoring cron jobs, AI mistakes in production and the real cost of the tools.
30 articles
A colleague spent three days getting GitHub Copilot running for five developers. Everyone's VPN config differs. Billing fails through certain payment methods. The corporate account needs a foreign card. After three days: two developers work with AI, three don't. The productivity gap shows up within a week.
I'm not tired of AI. I stopped talking to it.
Microsoft revoked enterprise Claude Code licenses for thousands of developers in a single procurement decision. One IT ticket, and the tool people had built their workflows around stopped working Monday morning.
I gave Claude Code a clear brief: refactor a PHP sync module in a Bitrix project. Type everything strictly. No mixed. No global state. All event handlers through explicit classes implementing a defined interface.
I spent an hour on an AI session only to discover at the end that the agent had rewritten a file we'd already fixed at the start. Not because it's broken — the context window filled up, and the model lost track of what happened early in the conversation. Since then I treat the context window like RAM: when it's full, you get swaps. So I built a session hygiene around that.
At the last project handoff, the client asked: "Can you show me where you wrote the code yourself and where the AI did it?"
The demo looked convincing. Eight agents, each closing its own ticket, the Kanban board moving like something out of a pitch deck. I'd been wanting to try something like it.
When people tell me "AI is going to replace developers," I think about one specific bug.
A year ago I wrote about what I don't delegate to AI. The list: architecture, cross-layer contracts, database schema decisions. That list still holds. But the year added something I didn't expect — a clearer picture of where Claude Code is genuinely useful.
When ChatGPT launched, the take I kept hearing was: juniors will code like midlevels, midlevels like seniors. The gap closes. Everyone levels up.
I added up my AI tool expenses for the year. $1,800. Did I save more than that? Honestly — I didn't know. So I sat down and calculated it.
Vibecoding is when you describe a task, the AI generates code, you check that it looks right, and you deploy. I don't work that way. Not because I'm afraid of AI — Claude Code is open on my screen every working day. It's because without structure, AI in a legacy PHP project isn't acceleration. It's debt with a nice interface.
Seven of my thirteen production tasks run through me. Five run without me. One I'll never hand to an algorithm — not even in "show me a draft first" mode.
I run 13 autonomous Claude tasks in my practice. They write LinkedIn posts, draft blog articles, run monitoring, send Telegram alerts. I trust them — and I sleep fine. Because if Claude writes something wrong in my name, I see it before it matters. But if it wrote that in response to a customer's question about a return policy on my client's e-commerce site? Different story. That's why I have a rule: no LLMs on client-facing features. Not yet.
When a client says "I want an AI agent to handle order processing," I ask one question: how much does processing one order cost right now?
AI code review tools are good at what they can see — syntax, patterns, obvious errors. They consistently miss what requires system context: cache dependencies, business logic hidden in refactors, and technical debt that only shows at scale.
I set up four MCP servers in one evening. Then removed two. Added one different one. Three months later I know what Model Context Protocol is: not a smarter AI. Less friction at specific handoff points.
A client asked me directly, three months ago: "You use Cursor and Claude Code — why should I pay your full rate?"
I set up Claude Code to write and publish posts on my blog. Every 20 minutes. No approval. Fully autonomous.
My practice runs 13 automations. They fire every day: collecting blog topics, generating candidates, publishing articles, sending notifications, logging everything.
I timed it. When I write a spec before opening Claude Code, an iteration takes around 40 minutes. When I skip straight to the prompt, it's closer to two and a half hours and three rollbacks.
The code looked correct. Tests passed. CI was green.
Claude wrote 340 lines of tests for me in eight minutes. Good tests. I'd have spent two hours on them. But when I asked it to suggest a module structure for a new Bitrix project, it produced something symmetric, logical, and completely wrong for our context. That gap became my working definition: where Claude's job ends and mine begins.
Last year I wrote roughly three times more code than the year before. That's true. AI tools genuinely accelerate. But a significant chunk of that code I later deleted.
I run 15 automated tasks. Five of them will never get auto-publish — even though I could technically remove the approval gate in an afternoon.
When I write a task for an agent that runs on a cron at 3am, I think about the wording differently.
I run 13 Claude agents on cron, fully unattended. Last week two of them failed. I found out within 2 minutes — not because I built a monitoring stack, but because I put five lines of JSONL into every task from day one.
When I handed 7 of 13 cron jobs to Claude, the first thing I built wasn't a prompt. It was a STOP file and an Inbox folder.
In one week my LinkedIn SSI dropped from 38 to 37. I'm fine with that. I just finished building a system of 13 cron jobs, and the first things I put in were a STOP file and an approval folder.
I run seven projects in parallel across four stacks. The docs for each one are current. Not because I'm tidy. Because I no longer write them.