Autonomous Development Supervisor

Built for Reliable, Continuous AI Software Engineering

LoopGoal keeps AI coding agents working toward a user-defined goal through small, verifiable, and atomic iterations.

The Core Loop

Observe → Select → Execute → Verify → Commit → Repeat

Most AI agents fail when asked to do unbounded tasks (e.g. “Refactor the entire codebase”). They make sweeping, broken changes across multiple files and hallucinate completion. LoopGoal solves this by enforcing bounded, single-improvement iterations validated by real test suites.

Core Engineering Principles

1. One Iteration, One Improvement

Each iteration focuses on exactly one bounded task. Never make unrelated sweeping changes in a single step.

2. Deterministic Verification Gates

Every iteration runs your project test commands (go test, pytest, npm test). If tests fail, changes are rejected and self-repaired.

3. Local Git Safety & Zero Silent Push

LoopGoal operates 100% locally. Remote push is disabled by default. Your existing uncommitted work is strictly protected.

4. Universal Agent Adapter Interface

Decoupled from AI providers. Works seamlessly with Antigravity, Claude Code, Cursor, and OpenAI Codex.

Architecture

cmd/loopgoal (CLI entry point)
      │
      ▼
   loop (Autonomous engine & state machine)
      │
 ┌────┼─────────┐
 ▼    ▼         ▼
agent git     verify (Deterministic test harness)
      │
      ▼
    state (.loopgoal/state.json)