How quickly product development with large language models under the hood is progressing
Last week, I spent several days painfully designing an MCP server that would serve as a unified access point for different agents and users to my knowledge base.
(“Normal” — is a subjective set of requirements that matter specifically to me).
Once I managed to consolidate the requirements into a complete design document that finally satisfied me, I launched the project with just three prompts in Codex.
First prompt
Me: Codex, here is RFC 01 — instructions on how to quickly start a project from scratch according to my requirements. Also add three small points I forgot in RFC (see about quick start).
Codex: Done.
Then I spent a couple more prompts to fix shortcomings and add details I hadn’t included in RFC.
Second prompt
Me: Codex, here is RFC 02 with the final design of the MCP server I created. Implement it.
Codex (after 15 minutes): No problem. Everything is ready.
I have already deployed the server and connected it to the current context. Now I asked it to test its operation in practice.
Codex (after 10 minutes): Almost everything works great, but there are a couple of points where I made a small mistake.
And then I experienced a real surprise — the agent immediately wrote MCP code and tested it without any issues in the same context. Moreover, it made several changes, tested them, applied them. Then played around with other modifications and rolled them back.
Third prompt (continuation)
Me: Forget about it. It doesn’t matter. Also, I asked another agent to upload my documents — it found a couple of issues and recorded them directly into the database, saying you could easily find them by the graph. Fix it.
Codex (after 5 minutes): Oh, that’s my mistake! Found the entry, reproduced the problem, and fixed it. Marked as completed. I can’t deploy it myself — your access is restricted, and I don’t have hands.
And here came a second surprise. Theoretically, I expected that the knowledge base — built on principles similar to OpenAI Engineering Harness and working through MCP (where instead of MD I have outlines and graphs, and instead of AGENTS.MD — manifests and so on) — would start working instantly. But how easily and quickly different agents connected and began interacting — that’s truly impressive. All thanks to the MCP protocol (yes, I admit defeat in the debate about it).
So, it turns out: after clearly defining my requirements for the agents, just an hour passed until:
1) the first version of MCP was launched;
2) Codex independently debugged it through the built-in MCP interface right in the same session;
3) various agents (Claude Desktop, Claude Cowork, Codex) connected seamlessly to the environment and started working.
Most of the time, I was occupied with completely different tasks.
Of course, this is still a prototype: graphs need to be added, agent behavior trajectories optimized, or a user interface developed — all of which will be tackled later.
But the main point is this: with a well-thought-out design document, anyone can create a working environment for collaborative work among multiple agents within an hour and start testing and using it immediately. Today, this was a real breakthrough for me: the bottleneck isn’t in development or technology — it’s in specific requirements and task formulations.
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby