The Agent Loop

The Agent Loop gives AI coding agents a stronger foundation to work from: clear project intent, one active goal, scoped memory, guardrails, verification, and handoff notes that survive past a single chat.

Most agents can write code. The hard part is keeping them from losing the plot after a few prompts: repeating failed fixes, drifting from the goal, over-editing unrelated files, or forgetting why a decision was made. The Agent Loop turns that messy process into a small project-local operating system for AI-assisted development.

It is just a .agent-loop/ folder you copy into a project. No service, daemon, account, or package manager required.

What It Adds

• A short onboarding flow that records what the project is trying to accomplish.
• A single active goal so the agent has one concrete outcome to finish before expanding scope.
• Project memory for decisions, observations, system maps, logs, and handoffs.
• Guardrails for risky changes, deleted code, dependencies, external effects, and stale context.
• Lightweight workflows for intake, planning, execution, safe deletion, and reflection.

The result is a steadier AI coding setup: fewer bug loops, clearer reviews, better resumability, and more stable code over time.

When To Use It

Use The Agent Loop when you want an AI assistant to work inside a real project, not just answer one-off coding questions.

It is especially useful when:

• multiple sessions need to build on the same project context;
• the codebase has rules, constraints, or risky areas the agent must remember;
• broad ideas need to become small, verifiable implementation goals;
• you want the agent to preserve useful lessons without dumping every transcript into future prompts.

Install v0.3

Download the uploaded v0.3 framework ZIP from GitHub Releases, not GitHub's automatic source-code ZIP. The framework ZIP contains only:

• .agent-loop/

Copy .agent-loop/ into the root of your project, then tell your coding agent:

Read `.agent-loop/START.md` and start The Agent Loop onboarding.

The onboarding flow inspects the repository, creates or carefully updates root AGENTS.md, asks the first blocking intake question, and records accepted project state before implementation begins.

How It Works

The framework keeps startup small and loads deeper context only when needed:

• .agent-loop/START.md defines what to load first and where to resume.
• .agent-loop/RULES.md defines modes, gates, scope rules, and context policy.
• .agent-loop/project/ stores intent, the active goal, roadmap, system map, observations, logs, and exceptions.
• .agent-loop/workflows/ stores procedures for intake, planning, execution, safe deletion, and reflection.
• .agent-loop/templates/ stores templates used only when creating or replacing matching state files.

This Repository

This repository is the source workspace for The Agent Loop. It contains release source folders, research notes, distilled case studies, and the project-local Agent Loop state used to develop the framework itself.

It is not meant to be cloned directly as a project framework install. Use the uploaded release ZIP asset instead.

Frozen release source folders live under releases/. The current public install target is releases/v0.3/.agent-loop/.