Self-Building Skills

When an agent hits a task no skill covers, it builds one: instantly available to every agent on your platform.

How it works

An agent receives a task.
It searches the existing skill library.
If nothing matches, it autonomously constructs the needed skill, using the built-in primitives as foundation components.
The new skill is saved to the library.
Every agent on the platform gains immediate access to it.

The 17 core primitives

Six categories, Communication, Code, Files, Memory, Data, and Media, cover capabilities ranging from Ask User and Web Search to Image Generation and Voice Transcription.

Compounding network effects

A sales agent building a "CRM enrichment" skill for its own use means every other agent on the platform can immediately reuse it. A support agent doing ticket enrichment doesn't need the same skill built twice. The library grows from an initial 17 primitives toward a projected 170, then 100,000+, as more agents encounter more novel tasks across an organization.

When a skill breaks

A failing skill isn't just left broken or silently deleted. Three paths exist depending on what changed:

Repair: an error traceback goes back through the same build pipeline that created the skill, which fixes it and re-tests before updating it in place.
Enhance: a feature request follows the same flow, adding capability without discarding what already worked.
Rollback: any prior version can be restored directly.

A skill that fails repeatedly is disabled automatically rather than left retrying forever; a successful repair re-enables it.

Building from an API spec

Beyond writing code from a plain-language description, a skill can also be built directly from a service's OpenAPI specification, which is often faster and more precise for a well-documented API than describing it in words.

Integrations: the external services skills call out to
Security: how self-built skills stay inside the same credential and egress boundaries as everything else