{"id":32186,"date":"2026-09-15T16:16:57","date_gmt":"2026-09-15T10:46:57","guid":{"rendered":"https:\/\/opstree.com\/blog\/?p=32186"},"modified":"2026-09-15T16:16:57","modified_gmt":"2026-09-15T10:46:57","slug":"mcp-agent-is-burning-tokens-before-it-even-starts","status":"publish","type":"post","link":"https:\/\/opstree.com\/blog\/mcp-agent-is-burning-tokens-before-it-even-starts\/","title":{"rendered":"Your MCP Agent Is Burning Tokens Before It Even Starts"},"content":{"rendered":"

A practical breakdown of Programmatic Tool Calling and Progressive Tool Discovery – two patterns that slash token overhead by up to 98% and make MCP-connected agents <\/span>actually scale<\/span>.<\/span><\/span>\u00a0<\/span><\/p>\n

The Invisible Tax on Every Agent Request<\/span><\/span><\/h2>\n

Imagine you hire a contractor to fix one leaky tap. But before they touch a single pipe, they unpack every tool from every truck, drills, welders, excavators and lay it all out in your kitchen. That’s\u00a0exactly what a naive MCP agent<\/a> does every single time it receives a user message.<\/span>\u00a0<\/span><\/p>\n

Here’s\u00a0what loading a typical 5-server MCP setup looks like\u00a0<\/span>before a single line of user content is processed:<\/span><\/b>\u00a0<\/span><\/p>\n

\n\n\n\n\n\n\n\n\n\n\n\n
MCP Server<\/th>\nTools<\/th>\nToken Cost<\/th>\n<\/tr>\n<\/thead>\n
GitHub<\/td>\n35 tools<\/td>\n~26,000 tokens<\/td>\n<\/tr>\n
Slack<\/td>\n11 tools<\/td>\n~21,000 tokens<\/td>\n<\/tr>\n
Jira<\/td>\n22 tools<\/td>\n~17,000 tokens<\/td>\n<\/tr>\n
Sentry<\/td>\n5 tools<\/td>\n~3,000 tokens<\/td>\n<\/tr>\n
Grafana<\/td>\n5 tools<\/td>\n~3,000 tokens<\/td>\n<\/tr>\n
Total<\/td>\n78 tools<\/td>\n~70,000 tokens<\/td>\n<\/tr>\n
With Tool Search<\/td>\nsame 78 tools<\/td>\n~8,700 tokens \u2705<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

Anthropic’s\u00a0own systems hit\u00a0<\/span>134,000 tokens<\/span><\/b>\u00a0in tool definitions alone before optimization. At that\u00a0point\u00a0your model\u00a0isn’t reasoning, it’s\u00a0drowning.<\/span>\u00a0<\/span><\/p>\n

And token cost\u00a0isn’t\u00a0even the only problem. The most common agent failures are\u00a0<\/span>wrong tool selection <\/span><\/b>and <\/span>incorrect parameters<\/span><\/b>, especially when tools have similar names like\u00a0<\/span>notification-send-user<\/span>\u00a0vs.\u00a0<\/span>notification-send-channel<\/span>. More tools in context = more confusion.<\/span>\u00a0<\/span><\/p>\n

Two Patterns That Fix This<\/span><\/span><\/h2>\n

Anthropic released three new beta features in November 2025 to address this:\u00a0<\/span><\/span>Tool Search Tool<\/span><\/span>,\u00a0<\/span><\/span>Programmatic Tool Calling<\/span><\/span>\u00a0and\u00a0<\/span><\/span>Tool Use Examples<\/span><\/span>. Together, they\u00a0<\/span>represent<\/span>\u00a0a fundamentally\u00a0<\/span>different way<\/span> of thinking about how agents interact with tools.<\/span><\/span><\/span><\/p>\n

Pattern 1: Progressive Tool Discovery (Tool Search Tool)<\/span><\/b>\u00a0<\/span><\/h3>\n

The core idea<\/span><\/b><\/h4>\n

Instead of loading all tool definitions upfront, tools are marked with\u00a0<\/span><\/span>defer_loading<\/span>: true<\/span><\/span>.\u00a0<\/span>They become <\/span><\/span>discoverable <\/span><\/span>without consuming tokens until actually needed.<\/span>\u00a0Claude only sees the Tool Search Tool itself (~500 tokens) at the start of each request.<\/span><\/span>\u00a0<\/span>\u00a0<\/span><\/p>\n

H<\/span><\/span><\/span>ow it works step by step<\/span><\/span>\u00a0<\/span><\/span><\/h4>\n

Step 1 – User sends a message, <\/span><\/b>Only\u00a0the Tool Search Tool (~500 tokens) is loaded in context. Nothing else.<\/span>\u00a0<\/span><\/p>\n

Step 2 – Agent searches for relevant tools, <\/span><\/b>Claude\u00a0queries\u00a0the search interface. Searching “github” returns only\u00a0<\/span>createPullRequest<\/span>\u00a0and\u00a0<\/span>listIssues<\/span> – not the other 33 GitHub tools, and certainly not your Slack, Jira, or Google Drive tools.<\/span>\u00a0<\/span><\/p>\n

Step 3 – Only matched tools expand into full definitions, <\/span><\/b>3\u20135 relevant tools (~3K tokens) are added to context. The rest\u00a0stay\u00a0dormant.<\/span>\u00a0<\/span><\/p>\n

Step 4 – Task executes with laser focus, <\/span><\/b>Total\u00a0context overhead: ~8,700 tokens vs. ~77,000 tokens.\u00a0<\/span>95% of context window\u00a0preserved.<\/span><\/b>\u00a0<\/span><\/p>\n

Implementation<\/span><\/span><\/span><\/h4>\n
\n
{\r\n  \"tools\": [\r\n    {\r\n      \"type\": \"tool_search_tool_regex_20251119\",\r\n      \"name\": \"tool_search_tool_regex\"\r\n    },\r\n    {\r\n      \"name\": \"github.createPullRequest\",\r\n      \"description\": \"Create a pull request\",\r\n      \"input_schema\": {},\r\n      \"defer_loading\": true\r\n    }\r\n  ]\r\n}<\/code><\/pre>\n<\/div>\n
For MCP servers, you can defer entire servers while keeping your highest-priority tools always loaded:<\/span><\/span>\u00a0<\/span><\/p>\n
\n
{\r\n  \"type\": \"mcp_toolset\",\r\n  \"mcp_server_name\": \"google-drive\",\r\n  \"default_config\": { \"defer_loading\": true },\r\n  \"configs\": {\r\n    \"search_files\": { \"defer_loading\": false }\r\n  }\r\n}<\/code><\/pre>\n<\/div>\n
The accuracy win<\/span><\/span><\/h4>\n
Beyond tokens, the\u00a0accuracy\u00a0improvement is striking. On internal MCP evaluations:<\/span>\u00a0<\/span><\/p>\n