Fix tense

Author: mnkiefer

scripts/paper/blog-post-draft.md

@@ -41,7 +41,7 @@ The single most common inefficiency the Optimizer identified was unused MCP tool
 
 When an agent connects to an MCP server, the server's entire tool manifest—every available function, with its full JSON schema—is included in the system prompt of every LLM API call for the duration of the session. A GitHub MCP server with 40 registered tools can add 10–15 KB of JSON schema to every turn's context. If the workflow only uses 5 of those tools, the other 35 are dead weight on every call.
 
-This is a very common pattern. Workflow authors naturally start with the full tool set available—it's the path of least resistance, and the agent can figure out which tools it needs. But as time goes on, most workflows settle into a narrow, stable set of tool calls. The Optimizer identifies this pattern by cross-referencing the tool manifest against the actual tool calls recorded in the MCP gateway logs and recommends pruning unused tools from the configuration.
+This is a very common pattern. Workflow authors naturally start with the full tool set available—it's the path of least resistance, and the agent can figure out which tools it needs. But as time goes on, most workflows settle into a narrow, stable set of tool calls. The Optimizer identified this pattern by cross-referencing the tool manifest against the actual tool calls recorded in the MCP gateway logs and recommends pruning unused tools from the configuration.
 
 For our smoke test workflows, removing unused tools from the MCP configuration reduced the system prompt by 8–12 KB per call, saving several thousand context tokens per run with no change to behavior.