Works

WORK 05 · Agent Harness

Superloop

Turns vibe coding from one reply into a budgeted execution loop.

A harnessed skill where the user acts as CEO and defines mission, scope, budget, and stop rule while the coding agent executes, verifies, records, and writes the next round's context back into runtime state.

SOURCE
Private GitHub repo
LIVE
No public live surface yet
superloop reelharness protocol
TYPE
Codex skill
STATE
JSON harness
CONTEXT
Runtime prompt
AUDIT
Evidence-gated stop

01

Problem

Many agent workflows fail because goals, gates, stop conditions, budgets, and unresolved gaps are not persistently tracked. Superloop moves that state out of the chat and into a harness, then renders reusable runtime context before the next round starts.

02

System

  • `SKILL.md` defines the operating contract and CEO / agent split.
  • `superloop_cli.sh` is the stable command entrypoint.
  • `src/superloop/cli.py` and the harness scripts own state, rounds, budget, and stop audits.
  • `context / next-prompt` renders mission, budget, active round, gap ledger, and completion audit into the next prompt.
  • `start-round` records hypothesis, change, round gate, and verification plan before implementation starts.
  • `record --mission-complete` requires completion evidence so the loop cannot stop on a claim alone.

03

Build trace

  • Installed and used as a Codex skill.
  • On 2026-05-13, runtime context, start-round, and completion evidence were merged into superloop main and pushed to GitHub.
  • This YoloLab project is using Superloop to record execution state and now documents the latest workflow.

04

Next path

Improve cross-tool state sync and failure recovery so Codex, Claude Code, and other agents can hand off from the same runtime context.

Loop / 2026