Anthropic is converging on code, not tool calls

Reading how Anthropic talks about building agents feels like watching a model converge during training.

They’re following an SGD-like trajectory toward process calling.

First, the MCP spec came out saying tool calls should be “focused and atomic.”

Then came Skills, where the advice shifted to “don’t just wrap APIs—build tools that enable complete tasks.”

And this month they released benchmarks showing that pairing LLMs with executable code cuts token use by 98.7%.

“Loops, conditionals, and error handling can be done with familiar code patterns rather than chaining individual tool calls.”

Tell me about it.

Waiting for an LLM to sample its way through a while loop is painful. It kills your latency and is prone to flaking out.

We’re finally moving past the “prompt and pray” phase.

Separation of concerns is back in style.