A live multi-agent loop where executors and reviewers take turns, exchange evidence, and keep improving over many sessions.
LOOP-STATION is a protocol skill for long-running agent work. It keeps Codex, Claude Code, and other agents coordinated when a task needs repeated execution, review, decisions, summaries, and next-session planning.
⭐ It is useful when progress depends on reading what happened, not only running a preset sweep. Unlike Hydra or grid-search sweeps that stay inside fixed settings, LOOP-STATION lets agents inspect code changes, logs, metrics, images, trends, and reviews, then actively decide what to try next.
Ask Codex:
Install the Codex skill from https://github.com/jjunsss/LOOP-STATION as loop-station.
Restart Codex after installation.
Ask Claude Code:
Install LOOP-STATION from https://github.com/jjunsss/LOOP-STATION.
Restart Claude Code after installation.
The loop is simple: one agent runs, another reviews, then the executor consumes that review and plans the next session.
Agents communicate through files under loop_station/, especially
loop_station/flags/. Each agent leaves its current state there, waits for the
other agent's turn to finish, then continues from the written artifacts.
Start from Codex with one $loop-station command. The format is flexible; the
important part is to brief the agents clearly for your project.
$loop-station
Goal: what you want to improve or decide.
Budget: sessions, time, GPUs/resources, or stop condition.
Evidence: metrics, logs, images, tests, or checks that should guide decisions.
Optional: reviewer role, paths, tools, ask-before rules, summaries.
Goal, Budget, and Evidence are recommended. Add anything else that helps control
your project. English/Korean Codex and Claude examples are here:
examples/quality-improvement-loop/.
I used LOOP-STATION once for a 50-session research-quality improvement run. In practice, one subject improved from roughly PSNR 24 to ~ 29 (PSNR is an explicit rendering quality metric) within about 10 hours. Codex kept running, Claude stayed in review standby through Monitor, and the run used far more tokens than I normally spend with a Pro subscription. The screenshots below are from that run.
- LOOP-STATION is a protocol skill, not an optimizer by itself. The executor still needs project-specific commands, metrics, artifacts, and resource checks.
- Give a session budget, resource budget, and stop condition before starting a long loop.
- Understand the permission level you give the agent.
- Give clear code-change rules: source edits, variants/backups, and ask-before rules.
- Claude Code reviewers can stay in standby with Monitor and act when the next review turn is ready.
- You can intervene mid-loop to redirect the next axis or stop a weak direction.