agent-loop-runtime

Agent Loop Runtime

Agent Loop Runtime is a portable, local-first self-check layer for coding agents and agentic apps.

It is built for one practical failure mode: an agent is about to act or make a strong claim, and it needs a cheap reminder to ask, verify, or stop before the mistake becomes expensive.

It is not a cloud service, a monitor, a policy oracle, or a general lie detector. It reads only the explicit text and files you give it.

Install

From PyPI:

pip install agent-loop-runtime

From a source checkout:

python -m agent_loop_runtime doctor --compact
python -m agent_loop_runtime release-check --compact

Quick use

agent-loop-runtime check --action "inspect this workspace read-only" --compact

Typical signals:

• PASS means continue within the stated scope.
• WARN means label uncertainty or verify before claiming.
• ASK means ask before expanding scope or taking a write/external/private action.
• BLOCK means do not proceed as stated.

Why it exists

Agent Loop Runtime turns a few useful habits into reusable infrastructure:

• authorization checks before actions;
• guess and evidence checks before strong claims;
• compact signals that humans and agents can both consume;
• local schemas and examples for repeatable integration.

The design goal is boring reliability: small checks, explicit boundaries, no unnecessary token burn, and no hidden network behavior.

Documentation

• Quickstart
• Runtime manual
• Threat model
• CI notes
• Packaging notes

Project

Mind Loops Lab is directed by one human owner, implemented with Codex, and reviewed with Claude as an independent second set of eyes.

License: MIT.