2026–present / Project
From a model response to controlled execution.
Local Agent Runner
The context
Local models need a defined interface to tools, a way to recover from supported mistakes, and a clear point at which execution stops. This Python runner supplies that execution layer.
My contribution
Model and tool integration
Connected Ollama to file, command, and Git tools in a containerized runner, with configurable tool and path permissions.
Parsing and repair
Implemented tool-call validation and supported argument coercion, returning repairable errors to the model within a bounded loop.
Run records
Retained tool transcripts and explicit completion, iteration-limit, and error states so callers can inspect a run beyond its final answer.
What came out of it
The public v2 delivery records an integration run through AI Project System. The implementation makes execution inspectable; a completed process state still requires checking the work it produced.