Agent Runtime
Agent Runtime is the bounded local task runner layered on top of Runtime Core.
It inspects a repository, asks Runtime Core for a plan and patch proposal, optionally materializes a file replacement, and can run one developer-supplied validation command.
Responsibilities
- inspect a local repository with a small tool surface
- request step-wise routing decisions from Runtime Core
- capture traces of selected SLMs, proposed changes, and validation results
- keep writes scoped to the repository root and controlled by
--writes
Inputs
- runtime bundle path
- local repository path
- task description
- optional validation command and trace output path
Outputs
- one JSON result describing the agent run
- optional trace file with routing and tool activity
- optional repository file replacement when writes are enabled
The no-model demo uses --dry-run so developers can inspect the end-to-end
agent flow without model downloads.