Essays

When Agents Build Their Own Tools

Agents no longer just call tools; they start forging their own. That rewrites our default assumptions about boundaries.

Tools become made, not called

这是内容管线的样例正文。

For the past two years, tool calling read as the end of the capability boundary. Now models write small, purpose-built tools for themselves, then chain those tools into one reproducible run. The engineering focus therefore moves from counting integrations to describing a problem well enough that the right tool gets built.

Rewriting the boundary assumptions

When tools are temporary and generated on demand, review shifts away from the stability of a single interface toward the constraints of generation: who approves it, how it is rolled back, and what evidence a failure leaves behind.

Leave traces that can be checked

Checkable traces beat elegant abstractions. The inputs, commands and exit codes of one build should be replayable by a third party after the fact.

Contents Edited by the PonyGo desk

KEY POINTS

  • 01 Tools become made, not called
  • 02 Rewriting the boundary assumptions
  • 03 Leave traces that can be checked