Buffaly builds the tool it needs, then uses it.
Most tool-using agents start with a fixed toolbox. When a required capability is missing, the task stops or gets handed back to a developer. This walkthrough shows Buffaly identifying the gap, creating the implementation, wiring it into the typed executable graph, and using it in the same body of work.
This is not a chatbot producing advice. It is a running agent extending its own executable capability graph.
Read the technical write-upThe capability is missing.
The running agent discovers the action graph does not yet contain the GitHub capability needed for the task.
Buffaly writes the implementation.
It creates supporting code and wires that behavior into a typed, callable action instead of stopping at advice.
The workflow keeps going.
The new action is activated and used in the same ongoing work without restarting the overall session.
What to watch for
- Buffaly uses normal language understanding to identify the missing capability, but execution remains bounded by typed actions and runtime state.
- The generated work becomes a reusable action in the capability graph, not a one-off chat response.
- The same environment preserves plans, evidence, tool calls, and validation so the work remains inspectable.
Try the runtime locally.
After watching the capability loop, start with the installer and quickstart to create your own prompt skill, ProtoScript skill, or DLL-backed action.