Save 98% on AI Agent Tokens With This One Trick

The title is misleading, this applies narrowly to MCP

Best summary ever. Did u miss the dead simple talk like a caveman hack? Programmatic tool calling seems good but hard to debug.

If you do not specify tools, mcp, skills ...... before session starts, the more AI model is "smarter" the less is chance that AI will use or even search for tools even when they are available. Maybe the best would be to use some free sub-agent as prompt encricher to gather skills, mcp, tools based on current prompt.

I do not see how secure it is. You can't set rules for code execution same way as for cli commands. Are you going to read all scripts ai writes?

This same topic pushed me to build a small MCP gateway a manoth ago around making standard MCP servers compatible with sandboxed code execution. It was inspired by the same articles you quoted. Curious what you think about this direction.

The problem is, it depends where you're using this, because a lot of tools either do a couple of things. You can inject the rules in via a proxy layer. What you're suggesting there though, I mean, that works fine, but if you're working in an agent environment, most agent environments will override the rule once it hits the tool. So you can only use this in specific cases.

Traditional software terms/technologies/artifacts were developed once and were used since ages ex API, json,yaml etc. but with AI problem seems to be never ending - needs context , lets call MCP, oh MCP is rotting - call tools, tools has too much info and is not giving right flexibility, lets invent skills, aah but the token is still burning lets invent toons. 90% problem of AI world is Claude is best, but it charges too much how to reduce the token spend and context saving so that we burn less money.

I hope you can feel the charm of this building.

the 98% token reduction trick is wild. been trying to figure out how to optimize my own agent workflows