Action ⟶ Dispatch ⟶ Execute

What I worked on

Bendover authored its first commit over the weekend. Nothing fancy. A tiny edit, but real.
As soon as I tried anything slightly non trivial, the loop broke down. It was behaving more like batch retries than an actual turn based agent
Spent time digging into how different systems structure tool calls and where the control actually lives

Every tool DSL feels messy. Either you invent your own and parse it, or the API invents one for you and you still parse it
I tried to escape that by using a strongly typed scripting surface with a clean SDK. Let the model emit executable code and skip text/JSON/context block glue
It is cleaner. The agent can call multiple tools sequentially without involving the orchestrator
But now I’m effectively running a programmable DSL. That means guards, sandboxes, safety rails
I’m also more dependant on the model behaving correctly. Otherwise errors will compound and I loose ability to inspect

All tool calling reduces to a constrained action language
If a model can choose external tools, there is always a dispatcher somewhere — for me it’s now the C# runtime
Action ⟶ dispatch ⟶ execute. You can hide it, but you can’t remove it

If I allow my programmable DSL to make multiple calls but return the CSX do I really loose observability?