Speaking at MCP Dev Summit Shanghai
Suvrakamal Das · Sep 10, 2026

In September I spoke at MCP Dev Summit Shanghai (AGNTCon + MCPCon China), the Linux Foundation's conference for the Model Context Protocol and agent tooling. It ran on September 6 and 7 at the Shanghai International Convention Center in Pudong, and I went on a scholarship to speak.

The talk
Jayita Bhattacharyya and I gave "How We Built Reliable and Self-Evolving Agent Skills" in the Agentic Engineering track, on the second day.

We started from a simple problem. Almost every layer of an agent runs on prompts: the skills, the tool descriptions, the MCP servers, the orchestration around them. Prompts are easy to write, but the outputs are unreliable, and tuning them by hand against a non-deterministic model is slow guesswork.
What we talked about instead is reflection. The agent looks at where it went wrong and rewrites its own instructions, so it gets better turn by turn, without RL or fine-tuning.
The hard part is knowing whether a change actually helped. So the loop needs evaluations and feedback built in, to measure which edit improved the whole system and which one didn't.

The slides and the session page are online.
Keep reading
Aug 24, 2026PyCon JP 2026, HiroshimaMy first PyCon in Japan, on a scholarship.
Jul 18, 2026My whole life story in a nutshellThe most reliable systems I have worked on were built from the dullest parts. That is not an accident.
May 30, 2026PyCon USA 2026My first ever PyCon
Jul 10, 2024Publication at SciPy Conference — accepted paperMy paper was accepted and published at SciPy 2024, Tacoma, WA, USA.