Context-shrinking tool cut AI agent token use 78% but lost two answers

A developer tested Hexum, a hosted MCP server that trims code before sending it to an AI model, against sending full files and against no files. Across 10 questions about open-source Python and TypeScript libraries, the trimmed approach cut tokens sent to the model by a median of 77.7%. Accuracy fell from 9 of 10 with full files to 7 of 10 with Hexum. Analysis of the misses showed the answer files were shown only by bare filename without their folder path, causing the agent to search wrong locations. The author also reported a bug in which repeated requests returned empty views, which the founder fixed within two days.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in