Skip to main content

Agent Runtime

Agent Runtime is the bounded local task runner layered on top of Runtime Core.

It inspects a repository, asks Runtime Core for routing and patch guidance, optionally writes changes, and can run one developer-supplied validation command.

What It Does​

  • inspects a local repository with a small tool surface
  • requests step-wise routing decisions from Runtime Core
  • captures traces of selected SLMs, proposed changes, and validation results
  • keeps writes scoped to the repository root and controlled by flags

What Goes In​

  • a runtime bundle or --slms-dir
  • a local repository path
  • a task description
  • optional validation and trace options

What Comes Out​

  • one JSON result describing the run
  • an optional trace file
  • optional repository changes when writes are enabled

Why It Matters​

This is the part most visitors care about: can the same local runtime actually do bounded coding work on a repository?

The answer in v0.1 is yes, but narrowly.

Current Boundaries​

  • local, single-run execution only
  • bounded tool loop, not a full IDE agent
  • no background orchestration or long-lived memory
  • dry-run validates control flow, not model quality