🔁 AgentLooper

Two AI agents. One builds. One reviews. Loop until it's right.

Claude Code writes the code. Codex reviews it. If there are issues, Claude Code fixes them. Repeat until approved. Zero config.

agentlooper "Add Stripe billing with usage-based pricing"

   ╭─────────────────────╮
   │   🔁 AgentLooper    │
   │   Multi-Agent CLI   │
   ╰─────────────────────╯

▶ Build (iteration 1)
  ⠼ Writing services/billing.ts (32s)
  ✔ Done

▶ Review (iteration 1)
  ⠧ Codex is reviewing... (18s)
  ⚠ Issues found
  │ Missing error handling in webhook endpoint
  │ No idempotency key on charge creation

▶ Fix (iteration 2)
  ⠹ Editing services/billing.ts (14s)
  ✔ Done

▶ Review (iteration 2)
  ✔ APPROVED

─────────────────────────────────────────────
  Status:     APPROVED
  Steps run:  4
  Duration:   124.3s
  Est. cost:  $0.28
─────────────────────────────────────────────

Why?

You've done this manually:

1. Claude Code builds a feature
2. You review it, find issues
3. You paste the feedback back
4. Repeat 3 more times

AgentLooper automates the entire loop. Two agents iterate until the code is right — you just walk away.

Install

npm install -g agentlooper

Prerequisites

You need both of these AI coding CLIs installed:

| Tool | Install | Role | |---|---|---| | Claude Code | npm install -g @anthropic-ai/claude-code | Builder — reads your codebase, writes and edits files | | Codex CLI | npm install -g @openai/codex | Reviewer — reviews the diff, finds bugs and issues |

Usage

cd your-project

agentlooper "Add user authentication with JWT tokens"

That's it. No config files. No YAML. No setup. Works with any tech stack — iOS, web, backend, CLI tools. The agents figure out what to do.

You can also pass the prompt without quotes:

agentlooper add a settings screen with dark mode toggle

More examples

# Build a feature
agentlooper "Add a /health endpoint that returns system status"

# Fix a bug
agentlooper "Fix the login bug where users get logged out after 5 minutes"

# Refactor
agentlooper "Refactor the database layer to use connection pooling"

# iOS
agentlooper "Add hidden contextual tags to items using Apple Intelligence"

How It Works

┌─────────────┐     ┌─────────────┐
│ Claude Code  │────▶│   Codex     │
│  (builder)   │     │ (reviewer)  │
│              │◀────│             │
│ Writes code  │     │ Reviews diff│
└─────────────┘     └─────────────┘
       │                    │
       └──── Loop until ────┘
              APPROVED

1. Build — Claude Code implements your request. It reads your codebase, writes files, runs commands — whatever is needed.
2. Review — Codex reviews all uncommitted changes (codex exec review --uncommitted --full-auto). It only reads the diff, never edits files.
3. If APPROVED → done.
4. If blocking issues found → Fix — Claude Code gets the feedback and fixes them → back to step 2.
5. Max 5 iterations, then exits with an error.

Approval logic

AgentLooper uses a three-tier system to decide whether the review passes:

| Signal | Action | |---|---| | A review line is exactly APPROVED | Approved immediately | | Only non-blocking feedback (nits, suggestions, style, docs, formatting) | Treated as approved — no fix loop | | Blocking issues (bugs, crashes, security, failing builds, data loss) | Triggers a fix iteration |

This prevents wasting iterations on cosmetic feedback while catching real issues.

Fail-fast

If any step (Build, Review, or Fix) exits with a non-zero code, AgentLooper stops immediately with an error. No silent failures.

Review output filtering

Codex prints session metadata (version, model, session ID, etc.) alongside its review. AgentLooper strips this noise so you only see the actual findings.

Cost tracking

Claude Code reports its API cost. AgentLooper accumulates the total across all iterations and shows it in the summary. Codex costs are not currently tracked.

What the spinner shows

While the agents work, the spinner shows real-time activity with an elapsed timer:

  ⠼ Reading Models/Item.swift (8s)
  ⠦ Writing Services/TagService.swift (24s)
  ⠧ Editing ContentView.swift (31s)
  ⠇ Running: xcodebuild -scheme App build (45s)
  ⠏ Searching files... (1m 12s)
  ⠙ Thinking... (1m 30s)

This is parsed from Claude Code's streaming JSON output — you see file reads, writes, edits, shell commands, searches, and thinking in real-time.

Configuration

None. AgentLooper is zero-config by design.

Under the hood, it runs:

• Claude Code: claude --print --dangerously-skip-permissions --verbose --output-format stream-json
• Codex: codex exec review --uncommitted --full-auto

Each command has a 20-minute timeout.

Exit codes

| Code | Meaning | |---|---| | 0 | Review approved (explicitly or no blocking issues) | | 1 | Build/Review/Fix failed, max iterations reached, or execution error |

Contributing

Contributions are welcome:

1. Report bugs — Open an issue with the error output
2. Improve review logic — Better blocking/non-blocking classification
3. Add features — PRs welcome

License

MIT