In January 2024 I built a side project called dAIlog. It was a multi-agent framework on top of AutoGen — a C# / ASP.NET front end talking to a Python service that ran seven agents in a GroupChat: a Planner, an Engineer who wrote Python, a Scientist who didn’t, a Coder, a Data analyst, a Critic who checked the others, and an Executor that ran the code.

A single user prompt would walk all the way through the loop and produce an artefact. I shot a fifteen-second screen recording to prove to myself it was real.

The prompt was “animate a realistic tesseract.” The video looks like this — silenced and laid sideways from the original phone recording:

dAIlog multi-agent system, January 2024 — a single prompt to draw a tesseract, the agent loop running, a 4D matplotlib projection at the end.
dAIlog, January 2024. Left: the agent loop running on a Python service. Right: the dAIlog chat UI talking to GPT-3.5 through the system. The matplotlib window at the end is the agent's output — a tesseract, 4D projected into 3D.

What happens in the recording, in order:

  1. The chat UI on the right shows the prompt going in.
  2. The IDE on the left shows the agent loop firing — agentchat.contrib.gpt_assistant_agent: Matching assistant found, using the first matching assistant.
  3. Some seconds pass. SpongeBob’s narrator appeared in the original — “two very boring minutes later” — though I have stripped him out of this version.
  4. A matplotlib window opens. Blue lines, cube-within-a-cube, the diagonal connections that make a four-dimensional hypercube legible when you flatten it into three.
  5. The Python file the system wrote is still on disk in the repo: dAIlog/dAiLogPythonService/paper/draw_line.py. It has not been touched since.

A few things worth saying about this, almost two years later.

One — this is what most people now call agentic coding and what the tooling industry has spent the last year of its life re-implementing. dAIlog was not unique; AutoGen itself was a published Microsoft project and plenty of people had similar things running. But I had it running, and I had built a proper UI in front of it, and I had cleaned the loop up to the point where the demo could be a single sentence.

Two — I proposed it as an internal product in early 2024. The proposal went nowhere, for the usual mix of reasons that proposals go nowhere inside a services company. Six months later Devin was funded, Cursor was on every developer’s machine, and Claude Code shipped. The lesson there is not really about being early; plenty of people were early. The lesson is about distribution. The thing you build matters less than the channel you stand on when you finish it.

Three — the point of keeping a fifteen-second video on your phone is not for anybody else. It is for the version of you who, two years from now, will need a reminder that the date on the file in the corner of the screen is real.

The Python file is still on disk. The agent loop, ported and modernised, became the basis of my QA Automation work. The repo is at github.com/msbel5/dAIlog for anyone who wants to look. SpongeBob has been retired with honour.