After my test drives with agentic development and different LMMs I wanted to have agentic measurement. I took remote connection to my Ubuntu machine running Claude Code, OpenCode and GitHub Copilot and split the terminal. I ran those all three agents simultaneously and gave them all the same prompt of checking the three different Copilot Studio agents starting with Föli Agent.
I was quite surprised that each of them voted Copilot Studio agent created by GitHub Copilot as the best solution, when looking at the source codes. I think Claude Code was the middle, then OpenCode performed the fastest and GitHub Copilot took the longest time.
What this experiment really reinforced for me is that agentic development is not just about speed, model preference, or brand. It is about how well an agent can understand intent, structure work, and stay consistent when dealing with real outputs over time.
In this setup, OpenCode felt clearly the fastest, Claude Code landed somewhere in the middle, and GitHub Copilot took the longest. Still, when all three agents reviewed the same Copilot Studio outputs, they independently voted the GitHub Copilot–created solution as the strongest from a source code perspective. That doesn’t make it “the best tool overall”, but it does make it the best fit for this particular task.
This was not meant to be a benchmark or a definitive ranking. Models change, tools evolve, and small prompt differences can shift results quickly. For me, the value was in observing agent behaviour — especially when multiple agents are solving the same real problem at the same time.
I will keep running these side‑by‑side experiments, especially as agentic workflows become more complex and long‑running. That is usually where the real differences start to show.
In the end, what matters most is not which agent talks the best — but which one actually does the work.
Action is still the most beautiful form of speech.